SCIENCETOSTARTUP
Bridging science and startup
2026-01-12 UTC | sha:12798ee9
Daily Research Map
Daily Snapshot
range [Jan 8]Filter by Tech Stack
Trending Today
LLM-Guided Quantified SMT Solving over Uninterpreted Functions ↗
Why It Matters
This solves the long-standing challenge of efficient SMT solving by reducing complexity through semantic understanding, which could improve automated reasoning tasks across many domains.
Science
Uses diffusion models to leverage large language models (LLMs) for generating candidate solutions, combined with traditional SMT techniques, offering a hybrid approach to improve solver performance.
Key Figures
Use Case Idea
Commercial usage can extend to enhanced solutions in industries that require advanced automated reasoning systems, such as hardware verification, software synthesis, and AI reasoning tasks.
Product Angle
How to sell this solution could involve targeting enterprises focused on developing or deploying automation in symbolic reasoning and decision-making systems, offering a specialized API service or integrated solver libraries.
Product Opportunity
Market size includes sectors like formal verification, AI development, and complex systems design, where improving solver performance is critical.
Disruption
Replaces traditional SMT solving methods with a more advanced approach that can handle previously intractable problem spaces.
Method & Eval
Tested on SMT-COMP benchmarks, demonstrating improved performance over existing state-of-the-art solvers like Z3 and CVC5, specifically in solving satisfiable formulas.
Caveats
Limitations include dependency on accurate LLM output and potential difficulties in maintaining soundness across diverse problem instances.
CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models ↗
Why It Matters
This research addresses a significant challenge in video-language models: the tendency to hallucinate actions and temporal sequences due to over-reliance on language priors. By generating counterfactual videos, the approach improves the models' ability to understand and reason about visual dynamics, leading to more accurate and reliable multimodal AI systems.
Science
The paper introduces a framework that uses multimodal large language models (LLMs) and diffusion-based video models to create counterfactual videos. These videos differ in actions or temporal structure while maintaining the same scene context, providing 'hard negatives' that help train video-language models to better understand and reason about visual dynamics.
Key Figures
Use Case Idea
Develop a tool for video content creators that automatically suggests edits to improve narrative coherence by identifying and correcting potential action or temporal hallucinations.
Product Angle
This research can be productized into a software tool or API that enhances existing video editing and analysis platforms by integrating counterfactual video generation capabilities to improve narrative accuracy and coherence.
Product Opportunity
The market for video content creation and analysis is vast, with applications in entertainment, education, and marketing. Companies and content creators would pay for tools that enhance video quality and accuracy, reducing the time and effort required for manual editing.
Disruption
This approach could replace existing video editing tools that rely heavily on manual input and language-based heuristics, offering a more automated and accurate solution for video content analysis.
Method & Eval
The framework was tested by building a synthetic dataset of ~26k preference pairs and fine-tuning a video-language model with a new optimization approach. The results showed consistent improvements in temporal ordering and effective transfer to standard video hallucination benchmarks.
Caveats
The approach relies on the quality of the generated counterfactual videos, which may not always perfectly mimic real-world scenarios. Additionally, the scalability and computational requirements of the framework could pose challenges for widespread adoption.
Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation ↗
Why It Matters
Accurate traffic data imputation is crucial for the development of intelligent transportation systems, which can lead to more efficient traffic management and reduced congestion. By improving imputation accuracy, FENCE can significantly enhance the reliability of traffic predictions and planning.
Science
The research introduces FENCE, a method that dynamically adjusts the guidance scale in diffusion models based on posterior likelihood approximations. This adaptive mechanism ensures that the imputation process remains closely aligned with observed data, even when observations are sparse, by leveraging spatial-temporal correlations.
Key Figures
Use Case Idea
Develop a traffic management platform that uses FENCE to provide real-time traffic updates and predictive analytics for urban planners and transportation agencies.
Product Angle
To productize this research, integrate FENCE into existing traffic management systems or develop a standalone platform that offers enhanced data imputation services for smart city initiatives.
Product Opportunity
The market for intelligent transportation systems is rapidly growing, driven by urbanization and the need for efficient traffic management. Transportation agencies and smart city planners are potential customers willing to pay for improved data accuracy and predictive capabilities.
Disruption
FENCE could replace less accurate imputation methods currently used in traffic management systems, offering a more reliable solution for handling missing data in traffic datasets.
Method & Eval
The method was tested on real-world traffic datasets, demonstrating significant improvements in imputation accuracy compared to existing methods. Key benchmarks include enhanced alignment with observed data and reduced drift toward prior distributions.
Caveats
The approach may require substantial computational resources and fine-tuning to adapt to different datasets and traffic conditions. Additionally, its performance in extremely sparse data scenarios remains to be fully validated.
Intelligence Feed(114)
LLM-Guided Quantified SMT Solving over Uninterpreted Functions
Quantified formulas with Uninterpreted Functions (UFs) over non-linear real arithmetic pose fundamental challenges for Satisfiability Modulo Theories (SMT) solving. Traditional quantifier instantiation methods struggle because they lack semantic understanding of UF constraints, forcing them to search through unbounded solution spaces with limited guidance. We present AquaForte, a framework that leverages Large Language Models to provide semantic guidance for UF instantiation by generating instantiated candidates for function definitions that satisfy the constraints, thereby significantly reducing the search space and complexity for solvers. Our approach preprocesses formulas through constraint separation, uses structured prompts to extract mathematical reasoning from LLMs, and integrates the results with traditional SMT algorithms through adaptive instantiation. AquaForte maintains soundness through systematic validation: LLM-guided instantiations yielding SAT solve the original problem, while UNSAT results generate exclusion clauses for iterative refinement. Completeness is preserved by fallback to traditional solvers augmented with learned constraints. Experimental evaluation on SMT-COMP benchmarks demonstrates that AquaForte solves numerous instances where state-of-the-art solvers like Z3 and CVC5 timeout, with particular effectiveness on satisfiable formulas. Our work shows that LLMs can provide valuable mathematical intuition for symbolic reasoning, establishing a new paradigm for SMT constraint solving.
CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models
Video-language models (VLMs) achieve strong multimodal understanding but remain prone to hallucinations, especially when reasoning about actions and temporal order. Existing mitigation strategies, such as textual filtering or random video perturbations, often fail to address the root cause: over-reliance on language priors rather than fine-grained visual dynamics. We propose a scalable framework for counterfactual video generation that synthesizes videos differing only in actions or temporal structure while preserving scene context. Our pipeline combines multimodal LLMs for action proposal and editing guidance with diffusion-based image and video models to generate semantic hard negatives at scale. Using this framework, we build CounterVid, a synthetic dataset of ~26k preference pairs targeting action recognition and temporal reasoning. We further introduce MixDPO, a unified Direct Preference Optimization approach that jointly leverages textual and visual preferences. Fine-tuning Qwen2.5-VL with MixDPO yields consistent improvements, notably in temporal ordering, and transfers effectively to standard video hallucination benchmarks. Code and models will be made publicly available.
Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation
Imputing missing values in spatial-temporal traffic data is essential for intelligent transportation systems. Among advanced imputation methods, score-based diffusion models have demonstrated competitive performance. These models generate data by reversing a noising process, using observed values as conditional guidance. However, existing diffusion models typically apply a uniform guidance scale across both spatial and temporal dimensions, which is inadequate for nodes with high missing data rates. Sparse observations provide insufficient conditional guidance, causing the generative process to drift toward the learned prior distribution rather than closely following the conditional observations, resulting in suboptimal imputation performance. To address this, we propose FENCE, a spatial-temporal feedback diffusion guidance method designed to adaptively control guidance scales during imputation. First, FENCE introduces a dynamic feedback mechanism that adjusts the guidance scale based on the posterior likelihood approximations. The guidance scale is increased when generated values diverge from observations and reduced when alignment improves, preventing overcorrection. Second, because alignment to observations varies across nodes and denoising steps, a global guidance scale for all nodes is suboptimal. FENCE computes guidance scales at the cluster level by grouping nodes based on their attention scores, leveraging spatial-temporal correlations to provide more accurate guidance. Experimental results on real-world traffic datasets show that FENCE significantly enhances imputation accuracy.
Orchestrating Intelligence: Confidence-Aware Routing for Efficient Multi-Agent Collaboration across Multi-Scale Models
While multi-agent systems (MAS) have demonstrated superior performance over single-agent approaches in complex reasoning tasks, they often suffer from significant computational inefficiencies. Existing frameworks typically deploy large language models (LLMs) uniformly across all agent roles, failing to account for the varying cognitive demands of different reasoning stages. We address this inefficiency by proposing OI-MAS framework, a novel multi-agent framework that implements an adaptive model-selection policy across a heterogeneous pool of multi-scale LLMs. Specifically, OI-MAS introduces a state-dependent routing mechanism that dynamically selects agent roles and model scales throughout the reasoning process. In addition, we introduce a confidence-aware mechanism that selects appropriate model scales conditioned on task complexity, thus reducing unnecessary reliance on large-scale models. Experimental results show that OI-MAS consistently outperforms baseline multi-agent systems, improving accuracy by up to 12.88\% while reducing cost by up to 79.78\%.
Atlas 2 -- Foundation models for clinical deployment
Pathology foundation models substantially advanced the possibilities in computational pathology -- yet tradeoffs in terms of performance, robustness, and computational requirements remained, which limited their clinical deployment. In this report, we present Atlas 2, Atlas 2-B, and Atlas 2-S, three pathology vision foundation models which bridge these shortcomings by showing state-of-the-art performance in prediction performance, robustness, and resource efficiency in a comprehensive evaluation across eighty public benchmarks. Our models were trained on the largest pathology foundation model dataset to date comprising 5.5 million histopathology whole slide images, collected from three medical institutions Charité - Universtätsmedizin Berlin, LMU Munich, and Mayo Clinic.
Code-Mix Sentiment Analysis on Hinglish Tweets
The effectiveness of brand monitoring in India is increasingly challenged by the rise of Hinglish--a hybrid of Hindi and English--used widely in user-generated content on platforms like Twitter. Traditional Natural Language Processing (NLP) models, built for monolingual data, often fail to interpret the syntactic and semantic complexity of this code-mixed language, resulting in inaccurate sentiment analysis and misleading market insights. To address this gap, we propose a high-performance sentiment classification framework specifically designed for Hinglish tweets. Our approach fine-tunes mBERT (Multilingual BERT), leveraging its multilingual capabilities to better understand the linguistic diversity of Indian social media. A key component of our methodology is the use of subword tokenization, which enables the model to effectively manage spelling variations, slang, and out-of-vocabulary terms common in Romanized Hinglish. This research delivers a production-ready AI solution for brand sentiment tracking and establishes a strong benchmark for multilingual NLP in low-resource, code-mixed environments.
FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via Mixture-of-Experts
Spatial-Temporal Graph (STG) forecasting on large-scale networks has garnered significant attention. However, existing models predominantly focus on short-horizon predictions and suffer from notorious computational costs and memory consumption when scaling to long-horizon predictions and large graphs. Targeting the above challenges, we present FaST, an effective and efficient framework based on heterogeneity-aware Mixture-of-Experts (MoEs) for long-horizon and large-scale STG forecasting, which unlocks one-week-ahead (672 steps at a 15-minute granularity) prediction with thousands of nodes. FaST is underpinned by two key innovations. First, an adaptive graph agent attention mechanism is proposed to alleviate the computational burden inherent in conventional graph convolution and self-attention modules when applied to large-scale graphs. Second, we propose a new parallel MoE module that replaces traditional feed-forward networks with Gated Linear Units (GLUs), enabling an efficient and scalable parallel structure. Extensive experiments on real-world datasets demonstrate that FaST not only delivers superior long-horizon predictive accuracy but also achieves remarkable computational efficiency compared to state-of-the-art baselines. Our source code is available at: https://github.com/yijizhao/FaST.
SpeechMedAssist: Efficiently and Effectively Adapting Speech Language Models for Medical Consultation
Medical consultations are intrinsically speech-centric. However, most prior works focus on long-text-based interactions, which are cumbersome and patient-unfriendly. Recent advances in speech language models (SpeechLMs) have enabled more natural speech-based interaction, yet the scarcity of medical speech data and the inefficiency of directly fine-tuning on speech data jointly hinder the adoption of SpeechLMs in medical consultation. In this paper, we propose SpeechMedAssist, a SpeechLM natively capable of conducting speech-based multi-turn interactions with patients. By exploiting the architectural properties of SpeechLMs, we decouple the conventional one-stage training into a two-stage paradigm consisting of (1) Knowledge & Capability Injection via Text and (2) Modality Re-alignment with Limited Speech Data, thereby reducing the requirement for medical speech data to only 10k synthesized samples. To evaluate SpeechLMs for medical consultation scenarios, we design a benchmark comprising both single-turn question answering and multi-turn simulated interactions. Experimental results show that our model outperforms all baselines in both effectiveness and robustness in most evaluation settings.
GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models
Multimodal Large Language Models (MLLMs) have demonstrated impressive progress in single-image grounding and general multi-image understanding. Recently, some methods begin to address multi-image grounding. However, they are constrained by single-target localization and limited types of practical tasks, due to the lack of unified modeling for generalized grounding tasks. Therefore, we propose GeM-VG, an MLLM capable of Generalized Multi-image Visual Grounding. To support this, we systematically categorize and organize existing multi-image grounding tasks according to their reliance of cross-image cues and reasoning, and introduce the MG-Data-240K dataset, addressing the limitations of existing datasets regarding target quantity and image relation. To tackle the challenges of robustly handling diverse multi-image grounding tasks, we further propose a hybrid reinforcement finetuning strategy that integrates chain-of-thought (CoT) reasoning and direct answering, considering their complementary strengths. This strategy adopts an R1-like algorithm guided by a carefully designed rule-based reward, effectively enhancing the model's overall perception and reasoning capabilities. Extensive experiments demonstrate the superior generalized grounding capabilities of our model. For multi-image grounding, it outperforms the previous leading MLLMs by 2.0% and 9.7% on MIG-Bench and MC-Bench, respectively. In single-image grounding, it achieves a 9.1% improvement over the base model on ODINW. Furthermore, our model retains strong capabilities in general multi-image understanding.
Categorical Belief Propagation: Sheaf-Theoretic Inference via Descent and Holonomy
We develop a categorical foundation for belief propagation on factor graphs. We construct the free hypergraph category \(\Syn_Σ\) on a typed signature and prove its universal property, yielding compositional semantics via a unique functor to the matrix category \(\cat{Mat}_R\). Message-passing is formulated using a Grothendieck fibration \(\int\Msg \to \cat{FG}_Σ\) over polarized factor graphs, with schedule-indexed endomorphisms defining BP updates. We characterize exact inference as effective descent: local beliefs form a descent datum when compatibility conditions hold on overlaps. This framework unifies tree exactness, junction tree algorithms, and loopy BP failures under sheaf-theoretic obstructions. We introduce HATCC (Holonomy-Aware Tree Compilation), an algorithm that detects descent obstructions via holonomy computation on the factor nerve, compiles non-trivial holonomy into mode variables, and reduces to tree BP on an augmented graph. Complexity is \(O(n^2 d_{\max} + c \cdot k_{\max} \cdot δ_{\max}^3 + n \cdot δ_{\max}^2)\) for \(n\) factors and \(c\) fundamental cycles. Experimental results demonstrate exact inference with significant speedup over junction trees on grid MRFs and random graphs, along with UNSAT detection on satisfiability instances.
CircuitLM: A Multi-Agent LLM-Aided Design Framework for Generating Circuit Schematics from Natural Language Prompts
Generating accurate circuit schematics from high-level natural language descriptions remains a persistent challenge in electronics design, as large language models (LLMs) frequently hallucinate in granular details, violate electrical constraints, and produce non-machine-readable outputs. We present CircuitLM, a novel multi-agent LLM-aided circuit design pipeline that translates user prompts into structured, visually interpretable CircuitJSON schematics through five sequential stages: (i) LLM-based component identification, (ii) canonical pinout retrieval, (iii) chain-of-thought reasoning by an electronics expert agent, (iv) JSON schematic synthesis, and (v) force-directed SVG visualization. Anchored by a curated, embedding-powered component knowledge base. While LLMs often violate electrical constraints, CircuitLM bridges this gap by grounding generation in a verified and dynamically extensible component database, initially comprising 50 components. To ensure safety, we incorporate a hybrid evaluation framework, namely Dual-Metric Circuit Validation (DMCV), validated against human-expert assessments, which achieves high fidelity in microcontroller-centric designs. We evaluate the system on 100 diverse embedded-systems prompts across six LLMs and introduce DMCV to assess both structural and electrical validity. This work bridges natural language input to deployable hardware designs, enabling reliable circuit prototyping by non-experts. Our code and data will be made public upon acceptance.
GUITester: Enabling GUI Agents for Exploratory Defect Discovery
Exploratory GUI testing is essential for software quality but suffers from high manual costs. While Multi-modal Large Language Model (MLLM) agents excel in navigation, they fail to autonomously discover defects due to two core challenges: \textit{Goal-Oriented Masking}, where agents prioritize task completion over reporting anomalies, and \textit{Execution-Bias Attribution}, where system defects are misidentified as agent errors. To address these, we first introduce \textbf{GUITestBench}, the first interactive benchmark for this task, featuring 143 tasks across 26 defects. We then propose \textbf{GUITester}, a multi-agent framework that decouples navigation from verification via two modules: (i) a \textit{Planning-Execution Module (PEM)} that proactively probes for defects via embedded testing intents, and (ii) a \textit{Hierarchical Reflection Module (HRM)} that resolves attribution ambiguity through interaction history analysis. GUITester achieves an F1-score of 48.90\% (Pass@3) on GUITestBench, outperforming state-of-the-art baselines (33.35\%). Our work demonstrates the feasibility of autonomous exploratory testing and provides a robust foundation for future GUI quality assurance~\footnote{Our code is now available in~\href{https://github.com/ADaM-BJTU/GUITestBench}{https://github.com/ADaM-BJTU/GUITestBench}}.
Vision-Language Agents for Interactive Forest Change Analysis
Modern forest monitoring workflows increasingly benefit from the growing availability of high-resolution satellite imagery and advances in deep learning. Two persistent challenges in this context are accurate pixel-level change detection and meaningful semantic change captioning for complex forest dynamics. While large language models (LLMs) are being adapted for interactive data exploration, their integration with vision-language models (VLMs) for remote sensing image change interpretation (RSICI) remains underexplored. To address this gap, we introduce an LLM-driven agent for integrated forest change analysis that supports natural language querying across multiple RSICI tasks. The proposed system builds upon a multi-level change interpretation (MCI) vision-language backbone with LLM-based orchestration. To facilitate adaptation and evaluation in forest environments, we further introduce the Forest-Change dataset, which comprises bi-temporal satellite imagery, pixel-level change masks, and multi-granularity semantic change captions generated using a combination of human annotation and rule-based methods. Experimental results show that the proposed system achieves mIoU and BLEU-4 scores of 67.10% and 40.17% on the Forest-Change dataset, and 88.13% and 34.41% on LEVIR-MCI-Trees, a tree-focused subset of LEVIR-MCI benchmark for joint change detection and captioning. These results highlight the potential of interactive, LLM-driven RSICI systems to improve accessibility, interpretability, and efficiency of forest change analysis. All data and code are publicly available at https://github.com/JamesBrockUoB/ForestChat.
A Closed-Loop Multi-Agent System Driven by LLMs for Meal-Level Personalized Nutrition Management
Personalized nutrition management aims to tailor dietary guidance to an individual's intake and phenotype, but most existing systems handle food logging, nutrient analysis and recommendation separately. We present a next-generation mobile nutrition assistant that combines image based meal logging with an LLM driven multi agent controller to provide meal level closed loop support. The system coordinates vision, dialogue and state management agents to estimate nutrients from photos and update a daily intake budget. It then adapts the next meal plan to user preferences and dietary constraints. Experiments with SNAPMe meal images and simulated users show competitive nutrient estimation, personalized menus and efficient task plans. These findings demonstrate the feasibility of multi agent LLM control for personalized nutrition and reveal open challenges in micronutrient estimation from images and in large scale real world studies.
SimuAgent: An LLM-Based Simulink Modeling Assistant Enhanced with Reinforcement Learning
Large language models (LLMs) have revolutionized text-based code automation, but their potential in graph-oriented engineering workflows remains under-explored. We introduce SimuAgent, an LLM-powered modeling and simulation agent tailored for Simulink. SimuAgent replaces verbose XML with a concise, dictionary-style Python representation, dramatically cutting token counts, improving interpretability, and enabling fast, in-process simulation. A lightweight plan-execute architecture, trained in two stages, equips the agent with both low-level tool skills and high-level design reasoning. To tackle sparse rewards in long-horizon tasks, we propose Reflection-GRPO (ReGRPO), which augments Group Relative Policy Optimization (GRPO) with self-reflection traces that supply rich intermediate feedback, accelerating convergence and boosting robustness. Experiments on SimuBench, our newly released benchmark comprising 5300 multi-domain modeling tasks, show that a Qwen2.5-7B model fine-tuned with SimuAgent converges faster and achieves higher modeling accuracy than standard RL baselines, and even surpasses GPT-4o when evaluated with few-shot prompting on the same benchmark. Ablations confirm that the two-stage curriculum and abstract-reconstruct data augmentation further enhance generalization. SimuAgent trains and runs entirely on-premise with modest hardware, delivering a privacy-preserving, cost-effective solution for industrial model-driven engineering. SimuAgent bridges the gap between LLMs and graphical modeling environments, offering a practical solution for AI-assisted engineering design in industrial settings.
GlimpRouter: Efficient Collaborative Inference by Glimpsing One Token of Thoughts
Large Reasoning Models (LRMs) achieve remarkable performance by explicitly generating multi-step chains of thought, but this capability incurs substantial inference latency and computational cost. Collaborative inference offers a promising solution by selectively allocating work between lightweight and large models, yet a fundamental challenge remains: determining when a reasoning step requires the capacity of a large model or the efficiency of a small model. Existing routing strategies either rely on local token probabilities or post-hoc verification, introducing significant inference overhead. In this work, we propose a novel perspective on step-wise collaboration: the difficulty of a reasoning step can be inferred from its very first token. Inspired by the "Aha Moment" phenomenon in LRMs, we show that the entropy of the initial token serves as a strong predictor of step difficulty. Building on this insight, we introduce GlimpRouter, a training-free step-wise collaboration framework. GlimpRouter employs a lightweight model to generate only the first token of each reasoning step and routes the step to a larger model only when the initial token entropy exceeds a threshold. Experiments on multiple benchmarks demonstrate that our approach significantly reduces inference latency while preserving accuracy. For instance, GlimpRouter attains a substantial 10.7% improvement in accuracy while reducing inference latency by 25.9% compared to a standalone large model on AIME25. These results suggest a simple yet effective mechanism for reasoning: allocating computation based on a glimpse of thought rather than full-step evaluation.
Self-MedRAG: a Self-Reflective Hybrid Retrieval-Augmented Generation Framework for Reliable Medical Question Answering
Large Language Models (LLMs) have demonstrated significant potential in medical Question Answering (QA), yet they remain prone to hallucinations and ungrounded reasoning, limiting their reliability in high-stakes clinical scenarios. While Retrieval-Augmented Generation (RAG) mitigates these issues by incorporating external knowledge, conventional single-shot retrieval often fails to resolve complex biomedical queries requiring multi-step inference. To address this, we propose Self-MedRAG, a self-reflective hybrid framework designed to mimic the iterative hypothesis-verification process of clinical reasoning. Self-MedRAG integrates a hybrid retrieval strategy, combining sparse (BM25) and dense (Contriever) retrievers via Reciprocal Rank Fusion (RRF) to maximize evidence coverage. It employs a generator to produce answers with supporting rationales, which are then assessed by a lightweight self-reflection module using Natural Language Inference (NLI) or LLM-based verification. If the rationale lacks sufficient evidentiary support, the system autonomously reformulates the query and iterates to refine the context. We evaluated Self-MedRAG on the MedQA and PubMedQA benchmarks. The results demonstrate that our hybrid retrieval approach significantly outperforms single-retriever baselines. Furthermore, the inclusion of the self-reflective loop yielded substantial gains, increasing accuracy on MedQA from 80.00% to 83.33% and on PubMedQA from 69.10% to 79.82%. These findings confirm that integrating hybrid retrieval with iterative, evidence-based self-reflection effectively reduces unsupported claims and enhances the clinical reliability of LLM-based systems.
Driving on Registers
We present DrivoR, a simple and efficient transformer-based architecture for end-to-end autonomous driving. Our approach builds on pretrained Vision Transformers (ViTs) and introduces camera-aware register tokens that compress multi-camera features into a compact scene representation, significantly reducing downstream computation without sacrificing accuracy. These tokens drive two lightweight transformer decoders that generate and then score candidate trajectories. The scoring decoder learns to mimic an oracle and predicts interpretable sub-scores representing aspects such as safety, comfort, and efficiency, enabling behavior-conditioned driving at inference. Despite its minimal design, DrivoR outperforms or matches strong contemporary baselines across NAVSIM-v1, NAVSIM-v2, and the photorealistic closed-loop HUGSIM benchmark. Our results show that a pure-transformer architecture, combined with targeted token compression, is sufficient for accurate, efficient, and adaptive end-to-end driving. Code and checkpoints will be made available via the project page.
Succeeding at Scale: Automated Multi-Retriever Fusion and Query-Side Adaptation for Multi-Tenant Search
Large-scale multi-tenant retrieval systems amass vast user query logs yet critically lack the curated relevance labels required for effective domain adaptation. This "dark data" problem is exacerbated by the operational cost of model updates: jointly fine-tuning query and document encoders requires re-indexing the entire corpus, which is prohibitive in multi-tenant environments with thousands of isolated indices. To address these dual challenges, we introduce \textbf{DevRev Search}, a passage retrieval benchmark for technical customer support constructed through a fully automatic pipeline. We employ a \textbf{fusion-based candidate generation} strategy, pooling results from diverse sparse and dense retrievers, and utilize an LLM-as-a-Judge to perform rigorous \textbf{consistency filtering} and relevance assignment. We further propose a practical \textbf{Index-Preserving Adaptation} strategy: by fine-tuning only the query encoder via Low-Rank Adaptation (LoRA), we achieve competitive performance improvements while keeping the document index frozen. Our experiments on DevRev Search and SciFact demonstrate that targeting specific transformer layers in the query encoder yields optimal quality-efficiency trade-offs, offering a scalable path for personalized enterprise search.
VERSE: Visual Embedding Reduction and Space Exploration. Clustering-Guided Insights for Training Data Enhancement in Visually-Rich Document Understanding
This work introduces VERSE, a methodology for analyzing and improving Vision-Language Models applied to Visually-rich Document Understanding by exploring their visual embedding space. VERSE enables the visualization of latent representations, supporting the assessment of model feasibility. It also facilitates the identification of problematic regions and guides the generation of synthetic data to enhance performance in those clusters. We validate the methodology by training on the synthetic MERIT Dataset and evaluating on its real-world counterpart, MERIT Secret. Results show that VERSE helps uncover the visual features associated with error-prone clusters, and that retraining with samples containing these features substantially boosts F1 performance without degrading generalization. Furthermore, we demonstrate that on-premise models such as Donut and Idefics2, when optimized with VERSE, match or even surpass the performance of SaaS solutions like GPT-4 and Pixtral.
RoboVIP: Multi-View Video Generation with Visual Identity Prompting Augments Robot Manipulation
The diversity, quantity, and quality of manipulation data are critical for training effective robot policies. However, due to hardware and physical setup constraints, collecting large-scale real-world manipulation data remains difficult to scale across diverse environments. Recent work uses text-prompt conditioned image diffusion models to augment manipulation data by altering the backgrounds and tabletop objects in the visual observations. However, these approaches often overlook the practical need for multi-view and temporally coherent observations required by state-of-the-art policy models. Further, text prompts alone cannot reliably specify the scene setup. To provide the diffusion model with explicit visual guidance, we introduce visual identity prompting, which supplies exemplar images as conditioning inputs to guide the generation of the desired scene setup. To this end, we also build a scalable pipeline to curate a visual identity pool from large robotics datasets. Using our augmented manipulation data to train downstream vision-language-action and visuomotor policy models yields consistent performance gains in both simulation and real-robot settings.
Miner:Mining Intrinsic Mastery for Data-Efficient RL in Large Reasoning Models
Current critic-free RL methods for large reasoning models suffer from severe inefficiency when training on positive homogeneous prompts (where all rollouts are correct), resulting in waste of rollouts due to zero advantage estimates. We introduce a radically simple yet powerful solution to \uline{M}ine \uline{in}trinsic mast\uline{er}y (Miner), that repurposes the policy's intrinsic uncertainty as a self-supervised reward signal, with no external supervision, auxiliary models, or additional inference cost. Our method pioneers two key innovations: (1) a token-level focal credit assignment mechanism that dynamically amplifies gradients on critical uncertain tokens while suppressing overconfident ones, and (2) adaptive advantage calibration to seamlessly integrate intrinsic and verifiable rewards. Evaluated across six reasoning benchmarks on Qwen3-4B and Qwen3-8B base models, Miner achieves state-of-the-art performance among the other four algorithms, yielding up to \textbf{4.58} absolute gains in Pass@1 and \textbf{6.66} gains in Pass@K compared to GRPO. Comparison with other methods targeted at exploration enhancement further discloses the superiority of the two newly proposed innovations. This demonstrates that latent uncertainty exploitation is both necessary and sufficient for efficient and scalable RL training of reasoning models.
TCAndon-Router: Adaptive Reasoning Router for Multi-Agent Collaboration
Multi-Agent Systems(MAS) have become a powerful paradigm for building high performance intelligent applications. Within these systems, the router responsible for determining which expert agents should handle a given query plays a crucial role in overall performance. Existing routing strategies generally fall into two categories: performance routing, which balances latency and cost across models of different sizes, and task routing, which assigns queries to domain-specific experts to improve accuracy. In real-world enterprise applications, task routing is more suitable; however, most existing approaches rely on static single-label decisions, which introduce two major limitations: (i) difficulty in seamlessly integrating new agents as business domains expand, and (ii) routing conflicts caused by overlapping agent capabilities, ultimately degrading accuracy and robustness.To address these challenges, we propose TCAndon-Router(TCAR): an adaptive reasoning router for multi-agent collaboration. Unlike traditional routers, TCAR supports dynamic agent onboarding and first generates a natural-language reasoning chain before predicting a set of candidate agents capable of handling the query. In addition, we design a collaborative execution pipeline in which selected agents independently produce responses, which are then aggregated and refined into a single high-quality response by a dedicated Refining Agent.Experiments on public datasets and real enterprise data demonstrate that TCAR significantly improves routing accuracy, reduces routing conflicts, and remains robust in ambiguous scenarios. We have released TCAR at https://huggingface.co/tencent/TCAndon-Router to support future research on explainable and collaborative multi-agent routing.
Orion-RAG: Path-Aligned Hybrid Retrieval for Graphless Data
Retrieval-Augmented Generation (RAG) has proven effective for knowledge synthesis, yet it encounters significant challenges in practical scenarios where data is inherently discrete and fragmented. In most environments, information is distributed across isolated files like reports and logs that lack explicit links. Standard search engines process files independently, ignoring the connections between them. Furthermore, manually building Knowledge Graphs is impractical for such vast data. To bridge this gap, we present Orion-RAG. Our core insight is simple yet effective: we do not need heavy algorithms to organize this data. Instead, we use a low-complexity strategy to extract lightweight paths that naturally link related concepts. We demonstrate that this streamlined approach suffices to transform fragmented documents into semi-structured data, enabling the system to link information across different files effectively. Extensive experiments demonstrate that Orion-RAG consistently outperforms mainstream frameworks across diverse domains, supporting real-time updates and explicit Human-in-the-Loop verification with high cost-efficiency. Experiments on FinanceBench demonstrate superior precision with a 25.2% relative improvement over strong baselines.
RelayLLM: Efficient Reasoning via Collaborative Decoding
Large Language Models (LLMs) for complex reasoning is often hindered by high computational costs and latency, while resource-efficient Small Language Models (SLMs) typically lack the necessary reasoning capacity. Existing collaborative approaches, such as cascading or routing, operate at a coarse granularity by offloading entire queries to LLMs, resulting in significant computational waste when the SLM is capable of handling the majority of reasoning steps. To address this, we propose RelayLLM, a novel framework for efficient reasoning via token-level collaborative decoding. Unlike routers, RelayLLM empowers the SLM to act as an active controller that dynamically invokes the LLM only for critical tokens via a special command, effectively "relaying" the generation process. We introduce a two-stage training framework, including warm-up and Group Relative Policy Optimization (GRPO) to teach the model to balance independence with strategic help-seeking. Empirical results across six benchmarks demonstrate that RelayLLM achieves an average accuracy of 49.52%, effectively bridging the performance gap between the two models. Notably, this is achieved by invoking the LLM for only 1.07% of the total generated tokens, offering a 98.2% cost reduction compared to performance-matched random routers.
Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human-Agent Interaction
As LLM-based agents are increasingly used in long-term interactions, cumulative memory is critical for enabling personalization and maintaining stylistic consistency. However, most existing systems adopt an ``all-or-nothing'' approach to memory usage: incorporating all relevant past information can lead to \textit{Memory Anchoring}, where the agent is trapped by past interactions, while excluding memory entirely results in under-utilization and the loss of important interaction history. We show that an agent's reliance on memory can be modeled as an explicit and user-controllable dimension. We first introduce a behavioral metric of memory dependence to quantify the influence of past interactions on current outputs. We then propose \textbf{Stee}rable \textbf{M}emory Agent, \texttt{SteeM}, a framework that allows users to dynamically regulate memory reliance, ranging from a fresh-start mode that promotes innovation to a high-fidelity mode that closely follows interaction history. Experiments across different scenarios demonstrate that our approach consistently outperforms conventional prompting and rigid memory masking strategies, yielding a more nuanced and effective control for personalized human-agent collaboration.
Token-Level LLM Collaboration via FusionRoute
Large language models (LLMs) exhibit strengths across diverse domains. However, achieving strong performance across these domains with a single general-purpose model typically requires scaling to sizes that are prohibitively expensive to train and deploy. On the other hand, while smaller domain-specialized models are much more efficient, they struggle to generalize beyond their training distributions. To address this dilemma, we propose FusionRoute, a robust and effective token-level multi-LLM collaboration framework in which a lightweight router simultaneously (i) selects the most suitable expert at each decoding step and (ii) contributes a complementary logit that refines or corrects the selected expert's next-token distribution via logit addition. Unlike existing token-level collaboration methods that rely solely on fixed expert outputs, we provide a theoretical analysis showing that pure expert-only routing is fundamentally limited: unless strong global coverage assumptions hold, it cannot in general realize the optimal decoding policy. By augmenting expert selection with a trainable complementary generator, FusionRoute expands the effective policy class and enables recovery of optimal value functions under mild conditions. Empirically, across both Llama-3 and Gemma-2 families and diverse benchmarks spanning mathematical reasoning, code generation, and instruction following, FusionRoute outperforms both sequence- and token-level collaboration, model merging, and direct fine-tuning, while remaining competitive with domain experts on their respective tasks.
Chain-of-Sanitized-Thoughts: Plugging PII Leakage in CoT of Large Reasoning Models
Large Reasoning Models (LRMs) improve performance, reliability, and interpretability by generating explicit chain-of-thought (CoT) reasoning, but this transparency introduces a serious privacy risk: intermediate reasoning often leaks personally identifiable information (PII) even when final answers are sanitized. We study how to induce privacy-first reasoning, where models reason without exposing sensitive information, using deployable interventions rather than post-hoc redaction. We introduce PII-CoT-Bench, a supervised dataset with privacy-aware CoT annotations, and a category-balanced evaluation benchmark covering realistic and adversarial leakage scenarios. Our results reveal a capability-dependent trend: state-of-the-art models benefit most from prompt-based controls, whereas weaker models require fine-tuning to achieve meaningful leakage reduction. Across models and categories, both approaches substantially reduce PII exposure with minimal degradation in utility, demonstrating that private reasoning can be achieved without sacrificing performance. Overall, we show that private CoT reasoning can be achieved with minimal utility loss, providing practical guidance for building privacy-preserving reasoning systems.
ConMax: Confidence-Maximizing Compression for Efficient Chain-of-Thought Reasoning
Recent breakthroughs in Large Reasoning Models (LRMs) have demonstrated that extensive Chain-of-Thought (CoT) generation is critical for enabling intricate cognitive behaviors, such as self-verification and backtracking, to solve complex tasks. However, this capability often leads to ``overthinking'', where models generate redundant reasoning paths that inflate computational costs without improving accuracy. While Supervised Fine-Tuning (SFT) on reasoning traces is a standard paradigm for the 'cold start' phase, applying existing compression techniques to these traces often compromises logical coherence or incurs prohibitive sampling costs. In this paper, we introduce ConMax (Confidence-Maximizing Compression), a novel reinforcement learning framework designed to automatically compress reasoning traces while preserving essential reasoning patterns. ConMax formulates compression as a reward-driven optimization problem, training a policy to prune redundancy by maximizing a weighted combination of answer confidence for predictive fidelity and thinking confidence for reasoning validity through a frozen auxiliary LRM. Extensive experiments across five reasoning datasets demonstrate that ConMax achieves a superior efficiency-performance trade-off. Specifically, it reduces inference length by 43% over strong baselines at the cost of a mere 0.7% dip in accuracy, proving its effectiveness in generating high-quality, efficient training data for LRMs.
Text as a Universal Interface for Transferable Personalization
We study the problem of personalization in large language models (LLMs). Prior work predominantly represents user preferences as implicit, model-specific vectors or parameters, yielding opaque ``black-box'' profiles that are difficult to interpret and transfer across models and tasks. In contrast, we advocate natural language as a universal, model- and task-agnostic interface for preference representation. The formulation leads to interpretable and reusable preference descriptions, while naturally supporting continual evolution as new interactions are observed. To learn such representations, we introduce a two-stage training framework that combines supervised fine-tuning on high-quality synthesized data with reinforcement learning to optimize long-term utility and cross-task transferability. Based on this framework, we develop AlignXplore+, a universal preference reasoning model that generates textual preference summaries. Experiments on nine benchmarks show that our 8B model achieves state-of-the-art performanc -- outperforming substantially larger open-source models -- while exhibiting strong transferability across tasks, model families, and interaction formats.
Precomputing Multi-Agent Path Replanning using Temporal Flexibility: A Case Study on the Dutch Railway Network
Executing a multi-agent plan can be challenging when an agent is delayed, because this typically creates conflicts with other agents. So, we need to quickly find a new safe plan. Replanning only the delayed agent often does not result in an efficient plan, and sometimes cannot even yield a feasible plan. On the other hand, replanning other agents may lead to a cascade of changes and delays. We show how to efficiently replan by tracking and using the temporal flexibility of other agents while avoiding cascading delays. This flexibility is the maximum delay an agent can take without changing the order of or further delaying more agents. Our algorithm, FlexSIPP, precomputes all possible plans for the delayed agent, also returning the changes for the other agents, for any single-agent delay within the given scenario. We demonstrate our method in a real-world case study of replanning trains in the densely-used Dutch railway network. Our experiments show that FlexSIPP provides effective solutions, relevant to real-world adjustments, and within a reasonable timeframe.
Higher-Order Knowledge Representations for Agentic Scientific Reasoning
Scientific inquiry requires systems-level reasoning that integrates heterogeneous experimental data, cross-domain knowledge, and mechanistic evidence into coherent explanations. While Large Language Models (LLMs) offer inferential capabilities, they often depend on retrieval-augmented contexts that lack structural depth. Traditional Knowledge Graphs (KGs) attempt to bridge this gap, yet their pairwise constraints fail to capture the irreducible higher-order interactions that govern emergent physical behavior. To address this, we introduce a methodology for constructing hypergraph-based knowledge representations that faithfully encode multi-entity relationships. Applied to a corpus of ~1,100 manuscripts on biocomposite scaffolds, our framework constructs a global hypergraph of 161,172 nodes and 320,201 hyperedges, revealing a scale-free topology (power law exponent ~1.23) organized around highly connected conceptual hubs. This representation prevents the combinatorial explosion typical of pairwise expansions and explicitly preserves the co-occurrence context of scientific formulations. We further demonstrate that equipping agentic systems with hypergraph traversal tools, specifically using node-intersection constraints, enables them to bridge semantically distant concepts. By exploiting these higher-order pathways, the system successfully generates grounded mechanistic hypotheses for novel composite materials, such as linking cerium oxide to PCL scaffolds via chitosan intermediates. This work establishes a "teacherless" agentic reasoning system where hypergraph topology acts as a verifiable guardrail, accelerating scientific discovery by uncovering relationships obscured by traditional graph methods.
Rethinking GNNs and Missing Features: Challenges, Evaluation and a Robust Solution
Handling missing node features is a key challenge for deploying Graph Neural Networks (GNNs) in real-world domains such as healthcare and sensor networks. Existing studies mostly address relatively benign scenarios, namely benchmark datasets with (a) high-dimensional but sparse node features and (b) incomplete data generated under Missing Completely At Random (MCAR) mechanisms. For (a), we theoretically prove that high sparsity substantially limits the information loss caused by missingness, making all models appear robust and preventing a meaningful comparison of their performance. To overcome this limitation, we introduce one synthetic and three real-world datasets with dense, semantically meaningful features. For (b), we move beyond MCAR and design evaluation protocols with more realistic missingness mechanisms. Moreover, we provide a theoretical background to state explicit assumptions on the missingness process and analyze their implications for different methods. Building on this analysis, we propose GNNmim, a simple yet effective baseline for node classification with incomplete feature data. Experiments show that GNNmim is competitive with respect to specialized architectures across diverse datasets and missingness regimes.
Defense Against Indirect Prompt Injection via Tool Result Parsing
As LLM agents transition from digital assistants to physical controllers in autonomous systems and robotics, they face an escalating threat from indirect prompt injection. By embedding adversarial instructions into the results of tool calls, attackers can hijack the agent's decision-making process to execute unauthorized actions. This vulnerability poses a significant risk as agents gain more direct control over physical environments. Existing defense mechanisms against Indirect Prompt Injection (IPI) generally fall into two categories. The first involves training dedicated detection models; however, this approach entails high computational overhead for both training and inference, and requires frequent updates to keep pace with evolving attack vectors. Alternatively, prompt-based methods leverage the inherent capabilities of LLMs to detect or ignore malicious instructions via prompt engineering. Despite their flexibility, most current prompt-based defenses suffer from high Attack Success Rates (ASR), demonstrating limited robustness against sophisticated injection attacks. In this paper, we propose a novel method that provides LLMs with precise data via tool result parsing while effectively filtering out injected malicious code. Our approach achieves competitive Utility under Attack (UA) while maintaining the lowest Attack Success Rate (ASR) to date, significantly outperforming existing methods. Code is available at GitHub.
APEX: Academic Poster Editing Agentic Expert
Designing academic posters is a labor-intensive process requiring the precise balance of high-density content and sophisticated layout. While existing paper-to-poster generation methods automate initial drafting, they are typically single-pass and non-interactive, often fail to align with complex, subjective user intent. To bridge this gap, we propose APEX (Academic Poster Editing agentic eXpert), the first agentic framework for interactive academic poster editing, supporting fine-grained control with robust multi-level API-based editing and a review-and-adjustment Mechanism. In addition, we introduce APEX-Bench, the first systematic benchmark comprising 514 academic poster editing instructions, categorized by a multi-dimensional taxonomy including operation type, difficulty, and abstraction level, constructed via reference-guided and reference-free strategies to ensure realism and diversity. We further establish a multi-dimensional VLM-as-a-judge evaluation protocol to assess instruction fulfillment, modification scope, and visual consistency & harmony. Experimental results demonstrate that APEX significantly outperforms baseline methods. Our implementation is available at https://github.com/Breesiu/APEX.
NC2C: Automated Convexification of Generic Non-Convex Optimization Problems
Non-convex optimization problems are pervasive across mathematical programming, engineering design, and scientific computing, often posing intractable challenges for traditional solvers due to their complex objective functions and constrained landscapes. To address the inefficiency of manual convexification and the over-reliance on expert knowledge, we propose NC2C, an LLM-based end-to-end automated framework designed to transform generic non-convex optimization problems into solvable convex forms using large language models. NC2C leverages LLMs' mathematical reasoning capabilities to autonomously detect non-convex components, select optimal convexification strategies, and generate rigorous convex equivalents. The framework integrates symbolic reasoning, adaptive transformation techniques, and iterative validation, equipped with error correction loops and feasibility domain correction mechanisms to ensure the robustness and validity of transformed problems. Experimental results on a diverse dataset of 100 generic non-convex problems demonstrate that NC2C achieves an 89.3\% execution rate and a 76\% success rate in producing feasible, high-quality convex transformations. This outperforms baseline methods by a significant margin, highlighting NC2C's ability to leverage LLMs for automated non-convex to convex transformation, reduce expert dependency, and enable efficient deployment of convex solvers for previously intractable optimization tasks.
SRU-Pix2Pix: A Fusion-Driven Generator Network for Medical Image Translation with Few-Shot Learning
Magnetic Resonance Imaging (MRI) provides detailed tissue information, but its clinical application is limited by long acquisition time, high cost, and restricted resolution. Image translation has recently gained attention as a strategy to address these limitations. Although Pix2Pix has been widely applied in medical image translation, its potential has not been fully explored. In this study, we propose an enhanced Pix2Pix framework that integrates Squeeze-and-Excitation Residual Networks (SEResNet) and U-Net++ to improve image generation quality and structural fidelity. SEResNet strengthens critical feature representation through channel attention, while U-Net++ enhances multi-scale feature fusion. A simplified PatchGAN discriminator further stabilizes training and refines local anatomical realism. Experimental results demonstrate that under few-shot conditions with fewer than 500 images, the proposed method achieves consistent structural fidelity and superior image quality across multiple intra-modality MRI translation tasks, showing strong generalization ability. These results suggest an effective extension of Pix2Pix for medical image translation.
Fast Mining and Dynamic Time-to-Event Prediction over Multi-sensor Data Streams
Given real-time sensor data streams obtained from machines, how can we continuously predict when a machine failure will occur? This work aims to continuously forecast the timing of future events by analyzing multi-sensor data streams. A key characteristic of real-world data streams is their dynamic nature, where the underlying patterns evolve over time. To address this, we present TimeCast, a dynamic prediction framework designed to adapt to these changes and provide accurate, real-time predictions of future event time. Our proposed method has the following properties: (a) Dynamic: it identifies the distinct time-evolving patterns (i.e., stages) and learns individual models for each, enabling us to make adaptive predictions based on pattern shifts. (b) Practical: it finds meaningful stages that capture time-varying interdependencies between multiple sensors and improve prediction performance; (c) Scalable: our algorithm scales linearly with the input size and enables online model updates on data streams. Extensive experiments on real datasets demonstrate that TimeCast provides higher prediction accuracy than state-of-the-art methods while finding dynamic changes in data streams with a great reduction in computational time.
RiskAtlas: Exposing Domain-Specific Risks in LLMs through Knowledge-Graph-Guided Harmful Prompt Generation
Large language models (LLMs) are increasingly applied in specialized domains such as finance and healthcare, where they introduce unique safety risks. Domain-specific datasets of harmful prompts remain scarce and still largely rely on manual construction; public datasets mainly focus on explicit harmful prompts, which modern LLM defenses can often detect and refuse. In contrast, implicit harmful prompts-expressed through indirect domain knowledge-are harder to detect and better reflect real-world threats. We identify two challenges: transforming domain knowledge into actionable constraints and increasing the implicitness of generated harmful prompts. To address them, we propose an end-to-end framework that first performs knowledge-graph-guided harmful prompt generation to systematically produce domain-relevant prompts, and then applies dual-path obfuscation rewriting to convert explicit harmful prompts into implicit variants via direct and context-enhanced rewriting. This framework yields high-quality datasets combining strong domain relevance with implicitness, enabling more realistic red-teaming and advancing LLM safety research. We release our code and datasets at GitHub.
TourPlanner: A Competitive Consensus Framework with Constraint-Gated Reinforcement Learning for Travel Planning
Travel planning is a sophisticated decision-making process that requires synthesizing multifaceted information to construct itineraries. However, existing travel planning approaches face several challenges: (1) Pruning candidate points of interest (POIs) while maintaining a high recall rate; (2) A single reasoning path restricts the exploration capability within the feasible solution space for travel planning; (3) Simultaneously optimizing hard constraints and soft constraints remains a significant difficulty. To address these challenges, we propose TourPlanner, a comprehensive framework featuring multi-path reasoning and constraint-gated reinforcement learning. Specifically, we first introduce a Personalized Recall and Spatial Optimization (PReSO) workflow to construct spatially-aware candidate POIs' set. Subsequently, we propose Competitive consensus Chain-of-Thought (CCoT), a multi-path reasoning paradigm that improves the ability of exploring the feasible solution space. To further refine the plan, we integrate a sigmoid-based gating mechanism into the reinforcement learning stage, which dynamically prioritizes soft-constraint satisfaction only after hard constraints are met. Experimental results on travel planning benchmarks demonstrate that TourPlanner achieves state-of-the-art performance, significantly surpassing existing methods in both feasibility and user-preference alignment.
ResMAS: Resilience Optimization in LLM-based Multi-agent Systems
Large Language Model-based Multi-Agent Systems (LLM-based MAS), where multiple LLM agents collaborate to solve complex tasks, have shown impressive performance in many areas. However, MAS are typically distributed across different devices or environments, making them vulnerable to perturbations such as agent failures. While existing works have studied the adversarial attacks and corresponding defense strategies, they mainly focus on reactively detecting and mitigating attacks after they occur rather than proactively designing inherently resilient systems. In this work, we study the resilience of LLM-based MAS under perturbations and find that both the communication topology and prompt design significantly influence system resilience. Motivated by these findings, we propose ResMAS: a two-stage framework for enhancing MAS resilience. First, we train a reward model to predict the MAS's resilience, based on which we train a topology generator to automatically design resilient topology for specific tasks through reinforcement learning. Second, we introduce a topology-aware prompt optimization method that refines each agent's prompt based on its connections and interactions with other agents. Extensive experiments across a range of tasks show that our approach substantially improves MAS resilience under various constraints. Moreover, our framework demonstrates strong generalization ability to new tasks and models, highlighting its potential for building resilient MASs.
Optimizing Path Planning using Deep Reinforcement Learning for UGVs in Precision Agriculture
This study focuses on optimizing path planning for unmanned ground vehicles (UGVs) in precision agriculture using deep reinforcement learning (DRL) techniques in continuous action spaces. The research begins with a review of traditional grid-based methods, such as A* and Dijkstra's algorithms, and discusses their limitations in dynamic agricultural environments, highlighting the need for adaptive learning strategies. The study then explores DRL approaches, including Deep Q-Networks (DQN), which demonstrate improved adaptability and performance in two-dimensional simulations. Enhancements such as Double Q-Networks and Dueling Networks are evaluated to further improve decision-making. Building on these results, the focus shifts to continuous action space models, specifically Deep Deterministic Policy Gradient (DDPG) and Twin Delayed Deep Deterministic Policy Gradient (TD3), which are tested in increasingly complex environments. Experiments conducted in a three-dimensional environment using ROS and Gazebo demonstrate the effectiveness of continuous DRL algorithms in navigating dynamic agricultural scenarios. Notably, the pretrained TD3 agent achieves a 95 percent success rate in dynamic environments, demonstrating the robustness of the proposed approach in handling moving obstacles while ensuring safety for both crops and the robot.
Know Thy Enemy: Securing LLMs Against Prompt Injection via Diverse Data Synthesis and Instruction-Level Chain-of-Thought Learning
Large language model (LLM)-integrated applications have become increasingly prevalent, yet face critical security vulnerabilities from prompt injection (PI) attacks. Defending against PI attacks faces two major issues: malicious instructions can be injected through diverse vectors, and injected instructions often lack clear semantic boundaries from the surrounding context, making them difficult to identify. To address these issues, we propose InstruCoT, a model enhancement method for PI defense that synthesizes diverse training data and employs instruction-level chain-of-thought fine-tuning, enabling LLMs to effectively identify and reject malicious instructions regardless of their source or position in the context. We evaluate InstruCoT across three critical dimensions: Behavior Deviation, Privacy Leakage, and Harmful Output. Experimental results across four LLMs demonstrate that InstruCoT significantly outperforms baselines in all dimensions while maintaining utility performance without degradation
Beyond the "Truth": Investigating Election Rumors on Truth Social During the 2024 Election
Large language models (LLMs) offer unprecedented opportunities for analyzing social phenomena at scale. This paper demonstrates the value of LLMs in psychological measurement by (1) compiling the first large-scale dataset of election rumors on a niche alt-tech platform, (2) developing a multistage Rumor Detection Agent that leverages LLMs for high-precision content classification, and (3) quantifying the psychological dynamics of rumor propagation, specifically the "illusory truth effect" in a naturalistic setting. The Rumor Detection Agent combines (i) a synthetic data-augmented, fine-tuned RoBERTa classifier, (ii) precision keyword filtering, and (iii) a two-pass LLM verification pipeline using GPT-4o mini. The findings reveal that sharing probability rises steadily with each additional exposure, providing large-scale empirical evidence for dose-response belief reinforcement in ideologically homogeneous networks. Simulation results further demonstrate rapid contagion effects: nearly one quarter of users become "infected" within just four propagation iterations. Taken together, these results illustrate how LLMs can transform psychological science by enabling the rigorous measurement of belief dynamics and misinformation spread in massive, real-world datasets.
FedKDX: Federated Learning with Negative Knowledge Distillation for Enhanced Healthcare AI Systems
This paper introduces FedKDX, a federated learning framework that addresses limitations in healthcare AI through Negative Knowledge Distillation (NKD). Unlike existing approaches that focus solely on positive knowledge transfer, FedKDX captures both target and non-target information to improve model generalization in healthcare applications. The framework integrates multiple knowledge transfer techniques--including traditional knowledge distillation, contrastive learning, and NKD--within a unified architecture that maintains privacy while reducing communication costs. Through experiments on healthcare datasets (SLEEP, UCI-HAR, and PAMAP2), FedKDX demonstrates improved accuracy (up to 2.53% over state-of-the-art methods), faster convergence, and better performance on non-IID data distributions. Theoretical analysis supports NKD's contribution to addressing statistical heterogeneity in distributed healthcare data. The approach shows promise for privacy-sensitive medical applications under regulatory frameworks like HIPAA and GDPR, offering a balanced solution between performance and practical implementation requirements in decentralized healthcare settings. The code and model are available at https://github.com/phamdinhdat-ai/Fed_2024.
Scaling Behavior Cloning Improves Causal Reasoning: An Open Model for Real-Time Video Game Playing
Behavior cloning is enjoying a resurgence in popularity as scaling both model and data sizes proves to provide a strong starting point for many tasks of interest. In this work, we introduce an open recipe for training a video game playing foundation model designed for inference in realtime on a consumer GPU. We release all data (8300+ hours of high quality human gameplay), training and inference code, and pretrained checkpoints under an open license. We show that our best model is capable of playing a variety of 3D video games at a level competitive with human play. We use this recipe to systematically examine the scaling laws of behavior cloning to understand how the model's performance and causal reasoning varies with model and data scale. We first show in a simple toy problem that, for some types of causal reasoning, increasing both the amount of training data and the depth of the network results in the model learning a more causal policy. We then systematically study how causality varies with the number of parameters (and depth) and training steps in scaled models of up to 1.2 billion parameters, and we find similar scaling results to what we observe in the toy problem.
Enhancing Multimodal Retrieval via Complementary Information Extraction and Alignment
Multimodal retrieval has emerged as a promising yet challenging research direction in recent years. Most existing studies in multimodal retrieval focus on capturing information in multimodal data that is similar to their paired texts, but often ignores the complementary information contained in multimodal data. In this study, we propose CIEA, a novel multimodal retrieval approach that employs Complementary Information Extraction and Alignment, which transforms both text and images in documents into a unified latent space and features a complementary information extractor designed to identify and preserve differences in the image representations. We optimize CIEA using two complementary contrastive losses to ensure semantic integrity and effectively capture the complementary information contained in images. Extensive experiments demonstrate the effectiveness of CIEA, which achieves significant improvements over both divide-and-conquer models and universal dense retrieval models. We provide an ablation study, further discussions, and case studies to highlight the advancements achieved by CIEA. To promote further research in the community, we have released the source code at https://github.com/zengdlong/CIEA.
BackdoorAgent: A Unified Framework for Backdoor Attacks on LLM-based Agents
Large language model (LLM) agents execute tasks through multi-step workflows that combine planning, memory, and tool use. While this design enables autonomy, it also expands the attack surface for backdoor threats. Backdoor triggers injected into specific stages of an agent workflow can persist through multiple intermediate states and adversely influence downstream outputs. However, existing studies remain fragmented and typically analyze individual attack vectors in isolation, leaving the cross-stage interaction and propagation of backdoor triggers poorly understood from an agent-centric perspective. To fill this gap, we propose \textbf{BackdoorAgent}, a modular and stage-aware framework that provides a unified, agent-centric view of backdoor threats in LLM agents. BackdoorAgent structures the attack surface into three functional stages of agentic workflows, including \textbf{planning attacks}, \textbf{memory attacks}, and \textbf{tool-use attacks}, and instruments agent execution to enable systematic analysis of trigger activation and propagation across different stages. Building on this framework, we construct a standardized benchmark spanning four representative agent applications: \textbf{Agent QA}, \textbf{Agent Code}, \textbf{Agent Web}, and \textbf{Agent Drive}, covering both language-only and multimodal settings. Our empirical analysis shows that \textit{triggers implanted at a single stage can persist across multiple steps and propagate through intermediate states.} For instance, when using a GPT-based backbone, we observe trigger persistence in 43.58\% of planning attacks, 77.97\% of memory attacks, and 60.28\% of tool-stage attacks, highlighting the vulnerabilities of the agentic workflow itself to backdoor threats. To facilitate reproducibility and future research, our code and benchmark are publicly available at GitHub.
A General Neural Backbone for Mixed-Integer Linear Optimization via Dual Attention
Mixed-integer linear programming (MILP), a widely used modeling framework for combinatorial optimization, are central to many scientific and engineering applications, yet remains computationally challenging at scale. Recent advances in deep learning address this challenge by representing MILP instances as variable-constraint bipartite graphs and applying graph neural networks (GNNs) to extract latent structural patterns and enhance solver efficiency. However, this architecture is inherently limited by the local-oriented mechanism, leading to restricted representation power and hindering neural approaches for MILP. Here we present an attention-driven neural architecture that learns expressive representations beyond the pure graph view. A dual-attention mechanism is designed to perform parallel self- and cross-attention over variables and constraints, enabling global information exchange and deeper representation learning. We apply this general backbone to various downstream tasks at the instance level, element level, and solving state level. Extensive experiments across widely used benchmarks show consistent improvements of our approach over state-of-the-art baselines, highlighting attention-based neural architectures as a powerful foundation for learning-enhanced mixed-integer linear optimization.
A Semi-supervised Molecular Learning Framework for Activity Cliff Estimation
Machine learning (ML) enables accurate and fast molecular property predictions, which are of interest in drug discovery and material design. Their success is based on the principle of similarity at its heart, assuming that similar molecules exhibit close properties. However, activity cliffs challenge this principle, and their presence leads to a sharp decline in the performance of existing ML algorithms, particularly graph-based methods. To overcome this obstacle under a low-data scenario, we propose a novel semi-supervised learning (SSL) method dubbed SemiMol, which employs predictions on numerous unannotated data as pseudo-signals for subsequent training. Specifically, we introduce an additional instructor model to evaluate the accuracy and trustworthiness of proxy labels because existing pseudo-labeling approaches require probabilistic outputs to reveal the model's confidence and fail to be applied in regression tasks. Moreover, we design a self-adaptive curriculum learning algorithm to progressively move the target model toward hard samples at a controllable pace. Extensive experiments on 30 activity cliff datasets demonstrate that SemiMol significantly enhances graph-based ML architectures and outpasses state-of-the-art pretraining and SSL baselines.
Specific Emitter Identification via Active Learning
With the rapid growth of wireless communications, specific emitter identification (SEI) is significant for communication security. However, its model training relies heavily on the large-scale labeled data, which are costly and time-consuming to obtain. To address this challenge, we propose an SEI approach enhanced by active learning (AL), which follows a three-stage semi-supervised training scheme. In the first stage, self-supervised contrastive learning is employed with a dynamic dictionary update mechanism to extract robust representations from large amounts of the unlabeled data. In the second stage, supervised training on a small labeled dataset is performed, where the contrastive and cross-entropy losses are jointly optimized to improve the feature separability and strengthen the classification boundaries. In the third stage, an AL module selects the most valuable samples from the unlabeled data for annotation based on the uncertainty and representativeness criteria, further enhancing generalization under limited labeling budgets. Experimental results on the ADS-B and WiFi datasets demonstrate that the proposed SEI approach significantly outperforms the conventional supervised and semi-supervised methods under limited annotation conditions, achieving higher recognition accuracy with lower labeling cost.
Scalable Floating-Point Satisfiability via Staged Optimization
This work introduces StageSAT, a new approach to solving floating-point satisfiability that bridges SMT solving with numerical optimization. StageSAT reframes a floating-point formula as a series of optimization problems in three stages of increasing precision. It begins with a fast, projection-aided descent objective to guide the search toward a feasible region, proceeding to bit-level accuracy with ULP$^2$ optimization and a final $n$-ULP lattice refinement. By construction, the final stage uses a representing function that is zero if and only if a candidate satisfies all constraints. Thus, when optimization drives the objective to zero, the resulting assignment is a valid solution, providing a built-in guarantee of soundness. To improve search, StageSAT introduces a partial monotone descent property on linear constraints via orthogonal projection, preventing the optimizer from stalling on flat or misleading landscapes. Critically, this solver requires no heavy bit-level reasoning or specialized abstractions; it treats complex arithmetic as a black-box, using runtime evaluations to navigate the input space. We implement StageSAT and evaluate it on extensive benchmarks, including SMT-COMP'25 suites and difficult cases from prior work. StageSAT proved more scalable and accurate than state-of-the-art optimization-based alternatives. It solved strictly more formulas than any competing solver under the same time budget, finding most satisfiable instances without producing spurious models. This amounts to 99.4% recall on satisfiable cases with 0% false SAT, exceeding the reliability of prior optimization-based solvers. StageSAT also delivered significant speedups (often 5--10$\times$) over traditional bit-precise SMT and numeric solvers. These results demonstrate that staged optimization significantly improves performance and correctness of floating-point satisfiability solving.
MineNPC-Task: Task Suite for Memory-Aware Minecraft Agents
We present \textsc{MineNPC-Task}, a user-authored benchmark and evaluation harness for testing memory-aware, mixed-initiative LLM agents in open-world \emph{Minecraft}. Rather than relying on synthetic prompts, tasks are elicited from formative and summative co-play with expert players, normalized into parametric templates with explicit preconditions and dependency structure, and paired with machine-checkable validators under a bounded-knowledge policy that forbids out-of-world shortcuts. The harness captures plan/act/memory events-including plan previews, targeted clarifications, memory reads and writes, precondition checks, and repair attempts and reports outcomes relative to the total number of attempted subtasks, derived from in-world evidence. As an initial snapshot, we instantiate the framework with GPT-4o and evaluate \textbf{216} subtasks across \textbf{8} experienced players. We observe recurring breakdown patterns in code execution, inventory/tool handling, referencing, and navigation, alongside recoveries supported by mixed-initiative clarifications and lightweight memory. Participants rated interaction quality and interface usability positively, while highlighting the need for stronger memory persistence across tasks. We release the complete task suite, validators, logs, and harness to support transparent, reproducible evaluation of future memory-aware embodied agents.
Internal Representations as Indicators of Hallucinations in Agent Tool Selection
Large Language Models (LLMs) have shown remarkable capabilities in tool calling and tool usage, but suffer from hallucinations where they choose incorrect tools, provide malformed parameters and exhibit 'tool bypass' behavior by performing simulations and generating outputs instead of invoking specialized tools or external systems. This undermines the reliability of LLM based agents in production systems as it leads to inconsistent results, and bypasses security and audit controls. Such hallucinations in agent tool selection require early detection and error handling. Unlike existing hallucination detection methods that require multiple forward passes or external validation, we present a computationally efficient framework that detects tool-calling hallucinations in real-time by leveraging LLMs' internal representations during the same forward pass used for generation. We evaluate this approach on reasoning tasks across multiple domains, demonstrating strong detection performance (up to 86.4\% accuracy) while maintaining real-time inference capabilities with minimal computational overhead, particularly excelling at detecting parameter-level hallucinations and inappropriate tool selections, critical for reliable agent deployment.
Evaluative Fingerprints: Stable and Systematic Differences in LLM Evaluator Behavior
LLM-as-judge systems promise scalable, consistent evaluation. We find the opposite: judges are consistent, but not with each other; they are consistent with themselves. Across 3,240 evaluations (9 judges x 120 unique video x pack items x 3 independent runs), inter-judge agreement is near-zero (Krippendorff's α = 0.042). On two dimensions, judges disagree more than random noise would predict (α < 0). Yet this disagreement isn't chaos; it's structured. A classifier identifies which judge produced an evaluation with 77.1% accuracy from rubric scores alone, rising to 89.9% with disposition features. Within model families, the signal is even stronger: GPT-4.1 and GPT-5.2 are distinguishable with 99.6% accuracy. We call this the reliability paradox: judges cannot agree on what constitutes quality, yet their disagreement patterns are so stable they function as fingerprints. Each judge implements a distinct, stable theory of quality: an "evaluative disposition" that shapes how it interprets any rubric. We characterize these dispositions along multiple axes: harshness/leniency, dimension emphasis, within-judge stability (ICC), and evidence behavior (receipt validity, semantic linkage via NLI, and shotgun index). The implication is stark: LLM judges are not interchangeable instruments measuring a shared construct. They are distinct measurement devices, each encoding its own implicit theory of quality. Averaging their scores produces a synthetic verdict that corresponds to no judge's actual values.
Reinforced Efficient Reasoning via Semantically Diverse Exploration
Reinforcement learning with verifiable rewards (RLVR) has proven effective in enhancing the reasoning of large language models (LLMs). Monte Carlo Tree Search (MCTS)-based extensions improve upon vanilla RLVR (e.g., GRPO) by providing tree-based reasoning rollouts that enable fine-grained and segment-level credit assignment. However, existing methods still suffer from limited exploration diversity and inefficient reasoning. To address the above challenges, we propose reinforced efficient reasoning via semantically diverse explorations, i.e., ROSE, for LLMs. To encourage more diverse reasoning exploration, our method incorporates a semantic-entropy-based branching strategy and an $\varepsilon$-exploration mechanism. The former operates on already sampled reasoning rollouts to capture semantic uncertainty and select branching points with high semantic divergence to generate new successive reasoning paths, whereas the latter stochastically initiates reasoning rollouts from the root, preventing the search process from becoming overly local. To improve efficiency, we design a length-aware segment-level advantage estimator that rewards concise and correct reasoning while penalizing unnecessarily long reasoning chains. Extensive experiments on various mathematical reasoning benchmarks with Qwen and Llama models validate the effectiveness and efficiency of ROSE. Codes are available at https://github.com/ZiqiZhao1/ROSE-rl.
Precision over Diversity: High-Precision Reward Generalizes to Robust Instruction Following
A central belief in scaling reinforcement learning with verifiable rewards for instruction following (IF) tasks is that, a diverse mixture of verifiable hard and unverifiable soft constraints is essential for generalizing to unseen instructions. In this work, we challenge this prevailing consensus through a systematic empirical investigation. Counter-intuitively, we find that models trained on hard-only constraints consistently outperform those trained on mixed datasets. Extensive experiments reveal that reward precision, rather than constraint diversity, is the primary driver of effective alignment. The LLM judge suffers from a low recall rate in detecting false response, which leads to severe reward hacking, thereby undermining the benefits of diversity. Furthermore, analysis of the attention mechanism reveals that high-precision rewards develop a transferable meta-skill for IF. Motivated by these insights, we propose a simple yet effective data-centric refinement strategy that prioritizes reward precision. Evaluated on five benchmarks, our approach outperforms competitive baselines by 13.4\% in performance while achieving a 58\% reduction in training time, maintaining strong generalization beyond instruction following. Our findings advocate for a paradigm shift: moving away from the indiscriminate pursuit of data diversity toward high-precision rewards.
Prototypicality Bias Reveals Blindspots in Multimodal Evaluation Metrics
Automatic metrics are now central to evaluating text-to-image models, often substituting for human judgment in benchmarking and large-scale filtering. However, it remains unclear whether these metrics truly prioritize semantic correctness or instead favor visually and socially prototypical images learned from biased data distributions. We identify and study \emph{prototypicality bias} as a systematic failure mode in multimodal evaluation. We introduce a controlled contrastive benchmark \textsc{\textbf{ProtoBias}} (\textit{\textbf{Proto}typical \textbf{Bias}}), spanning Animals, Objects, and Demography images, where semantically correct but non-prototypical images are paired with subtly incorrect yet prototypical adversarial counterparts. This setup enables a directional evaluation of whether metrics follow textual semantics or default to prototypes. Our results show that widely used metrics, including CLIPScore, PickScore, and VQA-based scores, frequently misrank these pairs, while even LLM-as-Judge systems exhibit uneven robustness in socially grounded cases. Human evaluations consistently favour semantic correctness with larger decision margins. Motivated by these findings, we propose \textbf{\textsc{ProtoScore}}, a robust 7B-parameter metric that substantially reduces failure rates and suppresses misranking, while running at orders of magnitude faster than the inference time of GPT-5, approaching the robustness of much larger closed-source judges.
AgentOCR: Reimagining Agent History via Optical Self-Compression
Recent advances in large language models (LLMs) enable agentic systems trained with reinforcement learning (RL) over multi-turn interaction trajectories, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token budgets and memory usage. We introduce AgentOCR, a framework that exploits the superior information density of visual tokens by representing the accumulated observation-action history as a compact rendered image. To make multi-turn rollouts scalable, AgentOCR proposes segment optical caching. By decomposing history into hashable segments and maintaining a visual cache, this mechanism eliminates redundant re-rendering. Beyond fixed rendering, AgentOCR introduces agentic self-compression, where the agent actively emits a compression rate and is trained with compression-aware reward to adaptively balance task success and token efficiency. We conduct extensive experiments on challenging agentic benchmarks, ALFWorld and search-based QA. Remarkably, results demonstrate that AgentOCR preserves over 95\% of text-based agent performance while substantially reducing token consumption (>50\%), yielding consistent token and memory efficiency. Our further analysis validates a 20x rendering speedup from segment optical caching and the effective strategic balancing of self-compression.
SciIF: Benchmarking Scientific Instruction Following Towards Rigorous Scientific Intelligence
As large language models (LLMs) transition from general knowledge retrieval to complex scientific discovery, their evaluation standards must also incorporate the rigorous norms of scientific inquiry. Existing benchmarks exhibit a critical blind spot: general instruction-following metrics focus on superficial formatting, while domain-specific scientific benchmarks assess only final-answer correctness, often rewarding models that arrive at the right result with the wrong reasons. To address this gap, we introduce scientific instruction following: the capability to solve problems while strictly adhering to the constraints that establish scientific validity. Specifically, we introduce SciIF, a multi-discipline benchmark that evaluates this capability by pairing university-level problems with a fixed catalog of constraints across three pillars: scientific conditions (e.g., boundary checks and assumptions), semantic stability (e.g., unit and symbol conventions), and specific processes(e.g., required numerical methods). Uniquely, SciIF emphasizes auditability, requiring models to provide explicit evidence of constraint satisfaction rather than implicit compliance. By measuring both solution correctness and multi-constraint adherence, SciIF enables finegrained diagnosis of compositional reasoning failures, ensuring that LLMs can function as reliable agents within the strict logical frameworks of science.
AT$^2$PO: Agentic Turn-based Policy Optimization via Tree Search
LLM agents have emerged as powerful systems for tackling multi-turn tasks by interleaving internal reasoning and external tool interactions. Agentic Reinforcement Learning has recently drawn significant research attention as a critical post-training paradigm to further refine these capabilities. In this paper, we present AT$^2$PO (Agentic Turn-based Policy Optimization via Tree Search), a unified framework for multi-turn agentic RL that addresses three core challenges: limited exploration diversity, sparse credit assignment, and misaligned policy optimization. AT$^2$PO introduces a turn-level tree structure that jointly enables Entropy-Guided Tree Expansion for strategic exploration and Turn-wise Credit Assignment for fine-grained reward propagation from sparse outcomes. Complementing this, we propose Agentic Turn-based Policy Optimization, a turn-level learning objective that aligns policy updates with the natural decision granularity of agentic interactions. ATPO is orthogonal to tree search and can be readily integrated into any multi-turn RL pipeline. Experiments across seven benchmarks demonstrate consistent improvements over the state-of-the-art baseline by up to 1.84 percentage points in average, with ablation studies validating the effectiveness of each component. Our code is available at https://github.com/zzfoutofspace/ATPO.
Smart IoT-Based Wearable Device for Detection and Monitoring of Common Cow Diseases Using a Novel Machine Learning Technique
Manual observation and monitoring of individual cows for disease detection present significant challenges in large-scale farming operations, as the process is labor-intensive, time-consuming, and prone to reduced accuracy. The reliance on human observation often leads to delays in identifying symptoms, as the sheer number of animals can hinder timely attention to each cow. Consequently, the accuracy and precision of disease detection are significantly compromised, potentially affecting animal health and overall farm productivity. Furthermore, organizing and managing human resources for the manual observation and monitoring of cow health is a complex and economically demanding task. It necessitates the involvement of skilled personnel, thereby contributing to elevated farm maintenance costs and operational inefficiencies. Therefore, the development of an automated, low-cost, and reliable smart system is essential to address these challenges effectively. Although several studies have been conducted in this domain, very few have simultaneously considered the detection of multiple common diseases with high prediction accuracy. However, advancements in Internet of Things (IoT), Machine Learning (ML), and Cyber-Physical Systems have enabled the automation of cow health monitoring with enhanced accuracy and reduced operational costs. This study proposes an IoT-enabled Cyber-Physical System framework designed to monitor the daily activities and health status of cow. A novel ML algorithm is proposed for the diagnosis of common cow diseases using collected physiological and behavioral data. The algorithm is designed to predict multiple diseases by analyzing a comprehensive set of recorded physiological and behavioral features, enabling accurate and efficient health assessment.
DSC2025 -- ViHallu Challenge: Detecting Hallucination in Vietnamese LLMs
The reliability of large language models (LLMs) in production environments remains significantly constrained by their propensity to generate hallucinations -- fluent, plausible-sounding outputs that contradict or fabricate information. While hallucination detection has recently emerged as a priority in English-centric benchmarks, low-to-medium resource languages such as Vietnamese remain inadequately covered by standardized evaluation frameworks. This paper introduces the DSC2025 ViHallu Challenge, the first large-scale shared task for detecting hallucinations in Vietnamese LLMs. We present the ViHallu dataset, comprising 10,000 annotated triplets of (context, prompt, response) samples systematically partitioned into three hallucination categories: no hallucination, intrinsic, and extrinsic hallucinations. The dataset incorporates three prompt types -- factual, noisy, and adversarial -- to stress-test model robustness. A total of 111 teams participated, with the best-performing system achieving a macro-F1 score of 84.80\%, compared to a baseline encoder-only score of 32.83\%, demonstrating that instruction-tuned LLMs with structured prompting and ensemble strategies substantially outperform generic architectures. However, the gap to perfect performance indicates that hallucination detection remains a challenging problem, particularly for intrinsic (contradiction-based) hallucinations. This work establishes a rigorous benchmark and explores a diverse range of detection methodologies, providing a foundation for future research into the trustworthiness and reliability of Vietnamese language AI systems.
Bridging Temporal and Textual Modalities: A Multimodal Framework for Automated Cloud Failure Root Cause Analysis
Root cause analysis in modern cloud infrastructure demands sophisticated understanding of heterogeneous data sources, particularly time-series performance metrics that involve core failure signatures. While large language models demonstrate remarkable capabilities in textual reasoning, their discrete token-based architecture creates fundamental incompatibilities with continuous numerical sequences exhibiting temporal dependencies. Current methodologies inadequately address this modality mismatch, constraining the potential of language model-driven automation in incident management workflows. This paper presents a multimodal diagnostic framework that harmonizes time-series representations with pretrained language model embedding spaces. Our approach contributes three technical advances: (1) a semantic compression technique that distills temporal segments into single-token abstractions while preserving pattern semantics, (2) an alignment encoder utilizing gated cross-attention to project time-series features into language model latent space, and (3) a retrieval-augmented diagnostic pipeline that synthesizes aligned embeddings with historical incident knowledge for expert-level failure attribution. Comprehensive evaluation across six cloud system benchmarks demonstrates that our framework achieves leading performance, reaching 48.75% diagnostic accuracy with notable improvements on scenarios involving compound failure modes. The results validate embedding-space alignment as an effective strategy for enabling language models to reason over multimodal telemetry data in production incident response contexts.
ToolGate: Contract-Grounded and Verified Tool Execution for LLMs
Large Language Models (LLMs) augmented with external tools have demonstrated remarkable capabilities in complex reasoning tasks. However, existing frameworks rely heavily on natural language reasoning to determine when tools can be invoked and whether their results should be committed, lacking formal guarantees for logical safety and verifiability. We present \textbf{ToolGate}, a forward execution framework that provides logical safety guarantees and verifiable state evolution for LLM tool calling. ToolGate maintains an explicit symbolic state space as a typed key-value mapping representing trusted world information throughout the reasoning process. Each tool is formalized as a Hoare-style contract consisting of a precondition and a postcondition, where the precondition gates tool invocation by checking whether the current state satisfies the required conditions, and the postcondition determines whether the tool's result can be committed to update the state through runtime verification. Our approach guarantees that the symbolic state evolves only through verified tool executions, preventing invalid or hallucinated results from corrupting the world representation. Experimental validation demonstrates that ToolGate significantly improves the reliability and verifiability of tool-augmented LLM systems while maintaining competitive performance on complex multi-step reasoning tasks. This work establishes a foundation for building more trustworthy and debuggable AI systems that integrate language models with external tools.
On the Limitations of Rank-One Model Editing in Answering Multi-hop Questions
Recent advances in Knowledge Editing (KE), particularly Rank-One Model Editing (ROME), show superior efficiency over fine-tuning and in-context learning for updating single-hop facts in transformers. However, these methods face significant challenges when applied to multi-hop reasoning tasks requiring knowledge chaining. In this work, we study the effect of editing knowledge with ROME on different layer depths and identify three key failure modes. First, the "hopping-too-late" problem occurs as later layers lack access to necessary intermediate representations. Second, generalization ability deteriorates sharply when editing later layers. Third, the model overfits to edited knowledge, incorrectly prioritizing edited-hop answers regardless of context. To mitigate the issues of "hopping-too-late" and generalisation decay, we propose Redundant Editing, a simple yet effective strategy that enhances multi-hop reasoning. Our experiments demonstrate that this approach can improve accuracy on 2-hop questions by at least 15.5 percentage points, representing a 96% increase over the previous single-edit strategy, while trading off some specificity and language naturalness.
Identifying Good and Bad Neurons for Task-Level Controllable LLMs
Large Language Models have demonstrated remarkable capabilities on multiple-choice question answering benchmarks, but the complex mechanisms underlying their large-scale neurons remain opaque, posing significant challenges for understanding and steering LLMs. While recent studies made progress on identifying responsible neurons for certain abilities, these ability-specific methods are infeasible for task-focused scenarios requiring coordinated use of multiple abilities. Moreover, these approaches focus only on supportive neurons that correlate positively with task completion, while neglecting neurons with other roles-such as inhibitive roles-and misled neuron attribution due to fortuitous behaviors in LLMs (i.e., correctly answer the questions by chance rather than genuine understanding). To address these challenges, we propose NeuronLLM, a novel task-level LLM understanding framework that adopts the biological principle of functional antagonism for LLM neuron identification. The key insight is that task performance is jointly determined by neurons with two opposing roles: good neurons that facilitate task completion and bad neurons that inhibit it. NeuronLLM achieves a holistic modeling of neurons via contrastive learning of good and bad neurons, while leveraging augmented question sets to mitigate the fortuitous behaviors in LLMs. Comprehensive experiments on LLMs of different sizes and families show the superiority of NeuronLLM over existing methods in four NLP tasks, providing new insights into LLM functional organization.
BanglaLorica: Design and Evaluation of a Robust Watermarking Algorithm for Large Language Models in Bangla Text Generation
As large language models (LLMs) are increasingly deployed for text generation, watermarking has become essential for authorship attribution, intellectual property protection, and misuse detection. While existing watermarking methods perform well in high-resource languages, their robustness in low-resource languages remains underexplored. This work presents the first systematic evaluation of state-of-the-art text watermarking methods: KGW, Exponential Sampling (EXP), and Waterfall, for Bangla LLM text generation under cross-lingual round-trip translation (RTT) attacks. Under benign conditions, KGW and EXP achieve high detection accuracy (>88%) with negligible perplexity and ROUGE degradation. However, RTT causes detection accuracy to collapse below RTT causes detection accuracy to collapse to 9-13%, indicating a fundamental failure of token-level watermarking. To address this, we propose a layered watermarking strategy that combines embedding-time and post-generation watermarks. Experimental results show that layered watermarking improves post-RTT detection accuracy by 25-35%, achieving 40-50% accuracy, representing a 3$\times$ to 4$\times$ relative improvement over single-layer methods, at the cost of controlled semantic degradation. Our findings quantify the robustness-quality trade-off in multilingual watermarking and establish layered watermarking as a practical, training-free solution for low-resource languages such as Bangla. Our code and data will be made public.
Decision-Aware Trust Signal Alignment for SOC Alert Triage
Detection systems that utilize machine learning are progressively implemented at Security Operations Centers (SOCs) to help an analyst to filter through high volumes of security alerts. Practically, such systems tend to reveal probabilistic results or confidence scores which are ill-calibrated and hard to read when under pressure. Qualitative and survey based studies of SOC practice done before reveal that poor alert quality and alert overload greatly augment the burden on the analyst, especially when tool outputs are not coherent with decision requirements, or signal noise. One of the most significant limitations is that model confidence is usually shown without expressing that there are asymmetric costs in decision making where false alarms are much less harmful than missed attacks. The present paper presents a decision-sensitive trust signal correspondence scheme of SOC alert triage. The framework combines confidence that has been calibrated, lightweight uncertainty cues, and cost-sensitive decision thresholds into coherent decision-support layer, instead of making changes to detection models. To enhance probabilistic consistency, the calibration is done using the known post-hoc methods and the uncertainty cues give conservative protection in situations where model certainty is low. To measure the model-independent performance of the suggested model, we apply the Logistic Regression and the Random Forest classifiers to the UNSW-NB15 intrusion detection benchmark. According to simulation findings, false negatives are greatly amplified by the presence of misaligned displays of confidence, whereas cost weighted loss decreases by orders of magnitude between models with decision aligned trust signals. Lastly, we describe a human-in-the-loop study plan that would allow empirically assessing the decision-making of the analysts with aligned and misaligned trust interfaces.
Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop
The rapid advancement of large language models (LLMs) has led to growing interest in using synthetic data to train future models. However, this creates a self-consuming retraining loop, where models are trained on their own outputs and may cause performance drops and induce emerging biases. In real-world applications, previously deployed LLMs may influence the data they generate, leading to a dynamic system driven by user feedback. For example, if a model continues to underserve users from a group, less query data will be collected from this particular demographic of users. In this study, we introduce the concept of \textbf{S}elf-\textbf{C}onsuming \textbf{P}erformative \textbf{L}oop (\textbf{SCPL}) and investigate the role of synthetic data in shaping bias during these dynamic iterative training processes under controlled performative feedback. This controlled setting is motivated by the inaccessibility of real-world user preference data from dynamic production systems, and enables us to isolate and analyze feedback-driven bias evolution in a principled manner. We focus on two types of loops, including the typical retraining setting and the incremental fine-tuning setting, which is largely underexplored. Through experiments on three real-world tasks, we find that the performative loop increases preference bias and decreases disparate bias. We design a reward-based rejection sampling strategy to mitigate the bias, moving towards more trustworthy self-improving systems.
Driver-Intention Prediction with Deep Learning: Real-Time Brain-to-Vehicle Communication
Brain-computer interfaces (BCIs) allow direct communication between the brain and electronics without the need for speech or physical movement. Such interfaces can be particularly beneficial in applications requiring rapid response times, such as driving, where a vehicle's advanced driving assistance systems could benefit from immediate understanding of a driver's intentions. This study presents a novel method for predicting a driver's intention to steer using electroencephalography (EEG) signals through deep learning. A driving simulator created a controlled environment in which participants imagined controlling a vehicle during various driving scenarios, including left and right turns, as well as straight driving. A convolutional neural network (CNN) classified the detected EEG data with minimal pre-processing. Our model achieved an accuracy of 83.7% in distinguishing between the three steering intentions and demonstrated the ability of CNNs to process raw EEG data effectively. The classification accuracy was highest for right-turn segments, which suggests a potential spatial bias in brain activity. This study lays the foundation for more intuitive brain-to-vehicle communication systems.
HMVI: Unifying Heterogeneous Attributes with Natural Neighbors for Missing Value Inference
Missing value imputation is a fundamental challenge in machine intelligence, heavily dependent on data completeness. Current imputation methods often handle numerical and categorical attributes independently, overlooking critical interdependencies among heterogeneous features. To address these limitations, we propose a novel imputation approach that explicitly models cross-type feature dependencies within a unified framework. Our method leverages both complete and incomplete instances to ensure accurate and consistent imputation in tabular data. Extensive experimental results demonstrate that the proposed approach achieves superior performance over existing techniques and significantly enhances downstream machine learning tasks, providing a robust solution for real-world systems with missing data.
An Empirical Investigation of Robustness in Large Language Models under Tabular Distortions
We investigate how large language models (LLMs) fail when tabular data in an otherwise canonical representation is subjected to semantic and structural distortions. Our findings reveal that LLMs lack an inherent ability to detect and correct subtle distortions in table representations. Only when provided with an explicit prior, via a system prompt, do models partially adjust their reasoning strategies and correct some distortions, though not consistently or completely. To study this phenomenon, we introduce a small, expert-curated dataset that explicitly evaluates LLMs on table question answering (TQA) tasks requiring an additional error-correction step prior to analysis. Our results reveal systematic differences in how LLMs ingest and interpret tabular information under distortion, with even SoTA models such as GPT-5.2 model exhibiting a drop of minimum 22% accuracy under distortion. These findings raise important questions for future research, particularly regarding when and how models should autonomously decide to realign tabular inputs, analogous to human behavior, without relying on explicit prompts or tabular data pre-processing.
AlgBench: To What Extent Do Large Reasoning Models Understand Algorithms?
Reasoning ability has become a central focus in the advancement of Large Reasoning Models (LRMs). Although notable progress has been achieved on several reasoning benchmarks such as MATH500 and LiveCodeBench, existing benchmarks for algorithmic reasoning remain limited, failing to answer a critical question: Do LRMs truly master algorithmic reasoning? To answer this question, we propose AlgBench, an expert-curated benchmark that evaluates LRMs under an algorithm-centric paradigm. AlgBench consists of over 3,000 original problems spanning 27 algorithms, constructed by ACM algorithmic experts and organized under a comprehensive taxonomy, including Euclidean-structured, non-Euclidean-structured, non-optimized, local-optimized, global-optimized, and heuristic-optimized categories. Empirical evaluations on leading LRMs (e.g., Gemini-3-Pro, DeepSeek-v3.2-Speciale and GPT-o3) reveal substantial performance heterogeneity: while models perform well on non-optimized tasks (up to 92%), accuracy drops sharply to around 49% on globally optimized algorithms such as dynamic programming. Further analysis uncovers \textbf{strategic over-shifts}, wherein models prematurely abandon correct algorithmic designs due to necessary low-entropy tokens. These findings expose fundamental limitations of problem-centric reinforcement learning and highlight the necessity of an algorithm-centric training paradigm for robust algorithmic reasoning.
T-Retriever: Tree-based Hierarchical Retrieval Augmented Generation for Textual Graphs
Retrieval-Augmented Generation (RAG) has significantly enhanced Large Language Models' ability to access external knowledge, yet current graph-based RAG approaches face two critical limitations in managing hierarchical information: they impose rigid layer-specific compression quotas that damage local graph structures, and they prioritize topological structure while neglecting semantic content. We introduce T-Retriever, a novel framework that reformulates attributed graph retrieval as tree-based retrieval using a semantic and structure-guided encoding tree. Our approach features two key innovations: (1) Adaptive Compression Encoding, which replaces artificial compression quotas with a global optimization strategy that preserves the graph's natural hierarchical organization, and (2) Semantic-Structural Entropy ($S^2$-Entropy), which jointly optimizes for both structural cohesion and semantic consistency when creating hierarchical partitions. Experiments across diverse graph reasoning benchmarks demonstrate that T-Retriever significantly outperforms state-of-the-art RAG methods, providing more coherent and contextually relevant responses to complex queries.
Key-Value Pair-Free Continual Learner via Task-Specific Prompt-Prototype
Continual learning aims to enable models to acquire new knowledge while retaining previously learned information. Prompt-based methods have shown remarkable performance in this domain; however, they typically rely on key-value pairing, which can introduce inter-task interference and hinder scalability. To overcome these limitations, we propose a novel approach employing task-specific Prompt-Prototype (ProP), thereby eliminating the need for key-value pairs. In our method, task-specific prompts facilitate more effective feature learning for the current task, while corresponding prototypes capture the representative features of the input. During inference, predictions are generated by binding each task-specific prompt with its associated prototype. Additionally, we introduce regularization constraints during prompt initialization to penalize excessively large values, thereby enhancing stability. Experiments on several widely used datasets demonstrate the effectiveness of the proposed method. In contrast to mainstream prompt-based approaches, our framework removes the dependency on key-value pairs, offering a fresh perspective for future continual learning research.
Belief in Authority: Impact of Authority in Multi-Agent Evaluation Framework
Multi-agent systems utilizing large language models often assign authoritative roles to improve performance, yet the impact of authority bias on agent interactions remains underexplored. We present the first systematic analysis of role-based authority bias in free-form multi-agent evaluation using ChatEval. Applying French and Raven's power-based theory, we classify authoritative roles into legitimate, referent, and expert types and analyze their influence across 12-turn conversations. Experiments with GPT-4o and DeepSeek R1 reveal that Expert and Referent power roles exert stronger influence than Legitimate power roles. Crucially, authority bias emerges not through active conformity by general agents, but through authoritative roles consistently maintaining their positions while general agents demonstrate flexibility. Furthermore, authority influence requires clear position statements, as neutral responses fail to generate bias. These findings provide key insights for designing multi-agent frameworks with asymmetric interaction patterns.
KnowMe-Bench: Benchmarking Person Understanding for Lifelong Digital Companions
Existing long-horizon memory benchmarks mostly use multi-turn dialogues or synthetic user histories, which makes retrieval performance an imperfect proxy for person understanding. We present \BenchName, a publicly releasable benchmark built from long-form autobiographical narratives, where actions, context, and inner thoughts provide dense evidence for inferring stable motivations and decision principles. \BenchName~reconstructs each narrative into a flashback-aware, time-anchored stream and evaluates models with evidence-linked questions spanning factual recall, subjective state attribution, and principle-level reasoning. Across diverse narrative sources, retrieval-augmented systems mainly improve factual accuracy, while errors persist on temporally grounded explanations and higher-level inferences, highlighting the need for memory mechanisms beyond retrieval. Our data is in \href{KnowMeBench}{https://github.com/QuantaAlpha/KnowMeBench}.
Estimating Causal Effects in Gaussian Linear SCMs with Finite Data
Estimating causal effects from observational data remains a fundamental challenge in causal inference, especially in the presence of latent confounders. This paper focuses on estimating causal effects in Gaussian Linear Structural Causal Models (GL-SCMs), which are widely used due to their analytical tractability. However, parameter estimation in GL-SCMs is often infeasible with finite data, primarily due to overparameterization. To address this, we introduce the class of Centralized Gaussian Linear SCMs (CGL-SCMs), a simplified yet expressive subclass where exogenous variables follow standardized distributions. We show that CGL-SCMs are equally expressive in terms of causal effect identifiability from observational distributions and present a novel EM-based estimation algorithm that can learn CGL-SCM parameters and estimate identifiable causal effects from finite observational samples. Our theoretical analysis is validated through experiments on synthetic data and benchmark causal graphs, demonstrating that the learned models accurately recover causal distributions.
Sci-Reasoning: A Dataset Decoding AI Innovation Patterns
While AI innovation accelerates rapidly, the intellectual process behind breakthroughs -- how researchers identify gaps, synthesize prior work, and generate insights -- remains poorly understood. The lack of structured data on scientific reasoning hinders systematic analysis and development of AI research agents. We introduce Sci-Reasoning, the first dataset capturing the intellectual synthesis behind high-quality AI research. Using community-validated quality signals and an LLM-accelerated, human-verified pipeline, we trace Oral and Spotlight papers across NeurIPS, ICML, and ICLR (2023-2025) to its key predecessors, articulating specific reasoning links in a structured format. Our analysis identifies 15 distinct thinking patterns, with three dominant strategies accounting for 52.7%: Gap-Driven Reframing (24.2%), Cross-Domain Synthesis (18.0%), and Representation Shift (10.5%). The most powerful innovation recipes combine multiple patterns: Gap-Driven Reframing + Representation Shift, Cross-Domain Synthesis + Representation Shift, and Gap-Driven Reframing + Cross-Domain Synthesis. This dataset enables quantitative studies of scientific progress and provides structured reasoning trajectories for training the next generation AI research agents.
Computational Compliance for AI Regulation: Blueprint for a New Research Domain
The era of AI regulation (AIR) is upon us. But AI systems, we argue, will not be able to comply with these regulations at the necessary speed and scale by continuing to rely on traditional, analogue methods of compliance. Instead, we posit that compliance with these regulations will only realistically be achieved computationally: that is, with algorithms that run across the life cycle of an AI system, automatically steering it toward AIR compliance in the face of dynamic conditions. Yet despite their (we would argue) inevitability, the research community has yet to specify exactly how these algorithms for computational AIR compliance should behave - or how we should benchmark their performance. To fill these gaps, we specify a set of design goals for such algorithms. In addition, we specify a benchmark dataset that can be used to quantitatively measure whether individual algorithms satisfy these design goals. By delivering this blueprint, we hope to give shape to an important but uncrystallized new domain of research - and, in doing so, incite necessary investment in it.
GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization
As language models become increasingly capable, users expect them to provide not only accurate responses but also behaviors aligned with diverse human preferences across a variety of scenarios. To achieve this, Reinforcement learning (RL) pipelines have begun incorporating multiple rewards, each capturing a distinct preference, to guide models toward these desired behaviors. However, recent work has defaulted to apply Group Relative Policy Optimization (GRPO) under multi-reward setting without examining its suitability. In this paper, we demonstrate that directly applying GRPO to normalize distinct rollout reward combinations causes them to collapse into identical advantage values, reducing the resolution of the training signal and resulting in suboptimal convergence and, in some cases, early training failure. We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues by decoupling the normalization of individual rewards, more faithfully preserving their relative differences and enabling more accurate multi-reward optimization, along with substantially improved training stability. We compare GDPO with GRPO across three tasks: tool calling, math reasoning, and coding reasoning, evaluating both correctness metrics (accuracy, bug ratio) and constraint adherence metrics (format, length). Across all settings, GDPO consistently outperforms GRPO, demonstrating its effectiveness and generalizability for multi-reward reinforcement learning optimization.
Arabic Prompts with English Tools: A Benchmark
Large Language Models (LLMs) are now integral to numerous industries, increasingly serving as the core reasoning engine for autonomous agents that perform complex tasks through tool-use. While the development of Arabic-native LLMs is accelerating, the benchmarks for evaluating their capabilities lag behind, with most existing frameworks focusing on English. A critical and overlooked area is tool-calling, where the performance of models prompted in non-English languages like Arabic is poorly understood, especially since these models are often pretrained on predominantly English data. This paper addresses this critical gap by introducing the first dedicated benchmark for evaluating the tool-calling and agentic capabilities of LLMs in the Arabic language. Our work provides a standardized framework to measure the functional accuracy and robustness of models in Arabic agentic workflows. Our findings reveal a huge performance gap: when users interact in Arabic, tool-calling accuracy drops by an average of 5-10\%, regardless of whether the tool descriptions themselves are in Arabic or English. By shedding light on these critical challenges, this benchmark aims to foster the development of more reliable and linguistically equitable AI agents for Arabic-speaking users.
Breaking Robustness Barriers in Cognitive Diagnosis: A One-Shot Neural Architecture Search Perspective
With the advancement of network technologies, intelligent tutoring systems (ITS) have emerged to deliver increasingly precise and tailored personalized learning services. Cognitive diagnosis (CD) has emerged as a core research task in ITS, aiming to infer learners' mastery of specific knowledge concepts by modeling the mapping between learning behavior data and knowledge states. However, existing research prioritizes model performance enhancement while neglecting the pervasive noise contamination in observed response data, significantly hindering practical deployment. Furthermore, current cognitive diagnosis models (CDMs) rely heavily on researchers' domain expertise for structural design, which fails to exhaustively explore architectural possibilities, thus leaving model architectures' full potential untapped. To address this issue, we propose OSCD, an evolutionary multi-objective One-Shot neural architecture search method for Cognitive Diagnosis, designed to efficiently and robustly improve the model's capability in assessing learner proficiency. Specifically, OSCD operates through two distinct stages: training and searching. During the training stage, we construct a search space encompassing diverse architectural combinations and train a weight-sharing supernet represented via the complete binary tree topology, enabling comprehensive exploration of potential architectures beyond manual design priors. In the searching stage, we formulate the optimal architecture search under heterogeneous noise scenarios as a multi-objective optimization problem (MOP), and develop an optimization framework integrating a Pareto-optimal solution search strategy with cross-scenario performance evaluation for resolution. Extensive experiments on real-world educational datasets validate the effectiveness and robustness of the optimal architectures discovered by our OSCD model for CD tasks.
Tape: A Cellular Automata Benchmark for Evaluating Rule-Shift Generalization in Reinforcement Learning
We present Tape, a controlled reinforcement-learning benchmark designed to isolate out-of-distribution (OOD) failure under latent rule shifts.Tape is derived from one-dimensional cellular automata, enabling precise train/test splits where observation and action spaces are held fixed while transition rules change. Using a reproducible evaluation pipeline, we compare model-free baselines, model-based planning with learned world models, and task-inference (meta-RL) methods. A consistent pattern emerges: methods that are strong in-distribution (ID) can collapse under heldout-rule OOD, and high-variance OOD evaluation can make rankings unstable unless experiments are sufficiently replicated.We provide (i) standardized OOD protocols, (ii) statistical reporting requirements (seeds, confidence intervals, and hypothesis tests), and (iii) information-theoretic identities connecting entropy reduction to conditional mutual information and expected posterior KL divergence, clarifying what "uncertainty reduction" objectives can and cannot guarantee under rule shifts.
AdaptEval: A Benchmark for Evaluating Large Language Models on Code Snippet Adaptation
Recent advancements in large language models (LLMs) have automated various software engineering tasks, with benchmarks emerging to evaluate their capabilities. However, for adaptation, a critical activity during code reuse, there is no benchmark to assess LLMs' performance, leaving their practical utility in this area unclear. To fill this gap, we propose AdaptEval, a benchmark designed to evaluate LLMs on code snippet adaptation. Unlike existing benchmarks, AdaptEval incorporates the following three distinctive features: First, Practical Context. Tasks in AdaptEval are derived from developers' practices, preserving rich contextual information from Stack Overflow and GitHub communities. Second, Multi-granularity Annotation. Each task is annotated with requirements at both task and adaptation levels, supporting the evaluation of LLMs across diverse adaptation scenarios. Third, Fine-grained Evaluation. AdaptEval includes a two-tier testing framework combining adaptation-level and function-level tests, which enables evaluating LLMs' performance across various individual adaptations. Based on AdaptEval, we conduct the first empirical study to evaluate six instruction-tuned LLMs and especially three reasoning LLMs on code snippet adaptation. Experimental results demonstrate that AdaptEval enables the assessment of LLMs' adaptation capabilities from various perspectives. It also provides critical insights into their current limitations, particularly their struggle to follow explicit instructions. We hope AdaptEval can facilitate further investigation and enhancement of LLMs' capabilities in code snippet adaptation, supporting their real-world applications.
Robust Reasoning as a Symmetry-Protected Topological Phase
Large language models suffer from "hallucinations"-logical inconsistencies induced by semantic noise. We propose that current architectures operate in a "Metric Phase," where causal order is vulnerable to spontaneous symmetry breaking. Here, we identify robust inference as an effective Symmetry-Protected Topological phase, where logical operations are formally isomorphic to non-Abelian anyon braiding, replacing fragile geometric interpolation with robust topological invariants. Empirically, we demonstrate a sharp topological phase transition: while Transformers and RNNs exhibit gapless decay, our Holonomic Network reveals a macroscopic "mass gap," maintaining invariant fidelity below a critical noise threshold. Furthermore, in a variable-binding task on $S_{10}$ ($3.6 \times 10^6$ states) representing symbolic manipulation, we demonstrate holonomic generalization: the topological model maintains perfect fidelity extrapolating $100\times$ beyond training ($L=50 \to 5000$), consistent with a theoretically indefinite causal horizon, whereas Transformers lose logical coherence. Ablation studies indicate this protection emerges strictly from non-Abelian gauge symmetry. This provides strong evidence for a new universality class for logical reasoning, linking causal stability to the topology of the semantic manifold.
Stock Market Price Prediction using Neural Prophet with Deep Neural Network
Stock market price prediction is a significant interdisciplinary research domain that depends at the intersection of finance, statistics, and economics. Forecasting Accurately predicting stock prices has always been a focal point for various researchers. However, existing statistical approaches for time-series prediction often fail to effectively forecast the probability range of future stock prices. Hence, to solve this problem, the Neural Prophet with a Deep Neural Network (NP-DNN) is proposed to predict stock market prices. The preprocessing technique used in this research is Z-score normalization, which normalizes stock price data by removing scale differences, making patterns easier to detect. Missing value imputation fills gaps in historical data, enhancing the models use of complete information for more accurate predictions. The Multi-Layer Perceptron (MLP) learns complex nonlinear relationships among stock market prices and extracts hidden patterns from the input data, thereby creating meaningful feature representations for better prediction accuracy. The proposed NP-DNN model achieved an accuracy of 99.21% compared with other approaches using the Fused Large Language Model. Keywords: deep neural network, forecasting stock prices, multi-layer perceptron, neural prophet, stock market price prediction.
Mechanisms of Prompt-Induced Hallucination in Vision-Language Models
Large vision-language models (VLMs) are highly capable, yet often hallucinate by favoring textual prompts over visual evidence. We study this failure mode in a controlled object-counting setting, where the prompt overstates the number of objects in the image (e.g., asking a model to describe four waterlilies when only three are present). At low object counts, models often correct the overestimation, but as the number of objects increases, they increasingly conform to the prompt regardless of the discrepancy. Through mechanistic analysis of three VLMs, we identify a small set of attention heads whose ablation substantially reduces prompt-induced hallucinations (PIH) by at least 40% without additional training. Across models, PIH-heads mediate prompt copying in model-specific ways. We characterize these differences and show that PIH ablation increases correction toward visual evidence. Our findings offer insights into the internal mechanisms driving prompt-induced hallucinations, revealing model-specific differences in how these behaviors are implemented.
CoV: Chain-of-View Prompting for Spatial Reasoning
Embodied question answering (EQA) in 3D environments often requires collecting context that is distributed across multiple viewpoints and partially occluded. However, most recent vision--language models (VLMs) are constrained to a fixed and finite set of input views, which limits their ability to acquire question-relevant context at inference time and hinders complex spatial reasoning. We propose Chain-of-View (CoV) prompting, a training-free, test-time reasoning framework that transforms a VLM into an active viewpoint reasoner through a coarse-to-fine exploration process. CoV first employs a View Selection agent to filter redundant frames and identify question-aligned anchor views. It then performs fine-grained view adjustment by interleaving iterative reasoning with discrete camera actions, obtaining new observations from the underlying 3D scene representation until sufficient context is gathered or a step budget is reached. We evaluate CoV on OpenEQA across four mainstream VLMs and obtain an average +11.56\% improvement in LLM-Match, with a maximum gain of +13.62\% on Qwen3-VL-Flash. CoV further exhibits test-time scaling: increasing the minimum action budget yields an additional +2.51\% average improvement, peaking at +3.73\% on Gemini-2.5-Flash. On ScanQA and SQA3D, CoV delivers strong performance (e.g., 116 CIDEr / 31.9 EM@1 on ScanQA and 51.1 EM@1 on SQA3D). Overall, these results suggest that question-aligned view selection coupled with open-view search is an effective, model-agnostic strategy for improving spatial reasoning in 3D EQA without additional training.
Vision-Language Introspection: Mitigating Overconfident Hallucinations in MLLMs via Interpretable Bi-Causal Steering
Object hallucination critically undermines the reliability of Multimodal Large Language Models, often stemming from a fundamental failure in cognitive introspection, where models blindly trust linguistic priors over specific visual evidence. Existing mitigations remain limited: contrastive decoding approaches operate superficially without rectifying internal semantic misalignments, while current latent steering methods rely on static vectors that lack instance-specific precision. We introduce Vision-Language Introspection (VLI), a training-free inference framework that simulates a metacognitive self-correction process. VLI first performs Attributive Introspection to diagnose hallucination risks via probabilistic conflict detection and localize the causal visual anchors. It then employs Interpretable Bi-Causal Steering to actively modulate the inference process, dynamically isolating visual evidence from background noise while neutralizing blind confidence through adaptive calibration. VLI achieves state-of-the-art performance on advanced models, reducing object hallucination rates by 12.67% on MMHal-Bench and improving accuracy by 5.8% on POPE.
Distilling the Thought, Watermarking the Answer: A Principle Semantic Guided Watermark for Large Reasoning Models
Reasoning Large Language Models (RLLMs) excelling in complex tasks present unique challenges for digital watermarking, as existing methods often disrupt logical coherence or incur high computational costs. Token-based watermarking techniques can corrupt the reasoning flow by applying pseudo-random biases, while semantic-aware approaches improve quality but introduce significant latency or require auxiliary models. This paper introduces ReasonMark, a novel watermarking framework specifically designed for reasoning-intensive LLMs. Our approach decouples generation into an undisturbed Thinking Phase and a watermarked Answering Phase. We propose a Criticality Score to identify semantically pivotal tokens from the reasoning trace, which are distilled into a Principal Semantic Vector (PSV). The PSV then guides a semantically-adaptive mechanism that modulates watermark strength based on token-PSV alignment, ensuring robustness without compromising logical integrity. Extensive experiments show ReasonMark surpasses state-of-the-art methods by reducing text Perplexity by 0.35, increasing translation BLEU score by 0.164, and raising mathematical accuracy by 0.67 points. These advancements are achieved alongside a 0.34% higher watermark detection AUC and stronger robustness to attacks, all with a negligible increase in latency. This work enables the traceable and trustworthy deployment of reasoning LLMs in real-world applications.
Agent-as-a-Judge
LLM-as-a-Judge has revolutionized AI evaluation by leveraging large language models for scalable assessments. However, as evaluands become increasingly complex, specialized, and multi-step, the reliability of LLM-as-a-Judge has become constrained by inherent biases, shallow single-pass reasoning, and the inability to verify assessments against real-world observations. This has catalyzed the transition to Agent-as-a-Judge, where agentic judges employ planning, tool-augmented verification, multi-agent collaboration, and persistent memory to enable more robust, verifiable, and nuanced evaluations. Despite the rapid proliferation of agentic evaluation systems, the field lacks a unified framework to navigate this shifting landscape. To bridge this gap, we present the first comprehensive survey tracing this evolution. Specifically, we identify key dimensions that characterize this paradigm shift and establish a developmental taxonomy. We organize core methodologies and survey applications across general and professional domains. Furthermore, we analyze frontier challenges and identify promising research directions, ultimately providing a clear roadmap for the next generation of agentic evaluation.
Compositional Steering of Large Language Models with Steering Tokens
Deploying LLMs in real-world applications requires controllable output that satisfies multiple desiderata at the same time. While existing work extensively addresses LLM steering for a single behavior, \textit{compositional steering} -- i.e., steering LLMs simultaneously towards multiple behaviors -- remains an underexplored problem. In this work, we propose \emph{compositional steering tokens} for multi-behavior steering. We first embed individual behaviors, expressed as natural language instructions, into dedicated tokens via self-distillation. Contrary to most prior work, which operates in the activation space, our behavior steers live in the space of input tokens, enabling more effective zero-shot composition. We then train a dedicated \textit{composition token} on pairs of behaviors and show that it successfully captures the notion of composition: it generalizes well to \textit{unseen} compositions, including those with unseen behaviors as well as those with an unseen \textit{number} of behaviors. Our experiments across different LLM architectures show that steering tokens lead to superior multi-behavior control compared to competing approaches (instructions, activation steering, and LoRA merging). Moreover, we show that steering tokens complement natural language instructions, with their combination resulting in further gains.
Challenges and Research Directions for Large Language Model Inference Hardware
Large Language Model (LLM) inference is hard. The autoregressive Decode phase of the underlying Transformer model makes LLM inference fundamentally different from training. Exacerbated by recent AI trends, the primary challenges are memory and interconnect rather than compute. To address these challenges, we highlight four architecture research opportunities: High Bandwidth Flash for 10X memory capacity with HBM-like bandwidth; Processing-Near-Memory and 3D memory-logic stacking for high memory bandwidth; and low-latency interconnect to speedup communication. While our focus is datacenter AI, we also review their applicability for mobile devices.
When to Act: Calibrated Confidence for Reliable Human Intention Prediction in Assistive Robotics
Assistive devices must determine both what a user intends to do and how reliable that prediction is before providing support. We introduce a safety-critical triggering framework based on calibrated probabilities for multimodal next-action prediction in Activities of Daily Living. Raw model confidence often fails to reflect true correctness, posing a safety risk. Post-hoc calibration aligns predicted confidence with empirical reliability and reduces miscalibration by about an order of magnitude without affecting accuracy. The calibrated confidence drives a simple ACT/HOLD rule that acts only when reliability is high and withholds assistance otherwise. This turns the confidence threshold into a quantitative safety parameter for assisted actions and enables verifiable behavior in an assistive control loop.
On the Definition and Detection of Cherry-Picking in Counterfactual Explanations
Counterfactual explanations are widely used to communicate how inputs must change for a model to alter its prediction. For a single instance, many valid counterfactuals can exist, which leaves open the possibility for an explanation provider to cherry-pick explanations that better suit a narrative of their choice, highlighting favourable behaviour and withholding examples that reveal problematic behaviour. We formally define cherry-picking for counterfactual explanations in terms of an admissible explanation space, specified by the generation procedure, and a utility function. We then study to what extent an external auditor can detect such manipulation. Considering three levels of access to the explanation process: full procedural access, partial procedural access, and explanation-only access, we show that detection is extremely limited in practice. Even with full procedural access, cherry-picked explanations can remain difficult to distinguish from non cherry-picked explanations, because the multiplicity of valid counterfactuals and flexibility in the explanation specification provide sufficient degrees of freedom to mask deliberate selection. Empirically, we demonstrate that this variability often exceeds the effect of cherry-picking on standard counterfactual quality metrics such as proximity, plausibility, and sparsity, making cherry-picked explanations statistically indistinguishable from baseline explanations. We argue that safeguards should therefore prioritise reproducibility, standardisation, and procedural constraints over post-hoc detection, and we provide recommendations for algorithm developers, explanation providers, and auditors.
What Students Ask, How a Generative AI Assistant Responds: Exploring Higher Education Students' Dialogues on Learning Analytics Feedback
Learning analytics dashboards (LADs) aim to support students' regulation of learning by translating complex data into feedback. Yet students, especially those with lower self-regulated learning (SRL) competence, often struggle to engage with and interpret analytics feedback. Conversational generative artificial intelligence (GenAI) assistants have shown potential to scaffold this process through real-time, personalised, dialogue-based support. Further advancing this potential, we explored authentic dialogues between students and GenAI assistant integrated into LAD during a 10-week semester. The analysis focused on questions students with different SRL levels posed, the relevance and quality of the assistant's answers, and how students perceived the assistant's role in their learning. Findings revealed distinct query patterns. While low SRL students sought clarification and reassurance, high SRL students queried technical aspects and requested personalised strategies. The assistant provided clear and reliable explanations but limited in personalisation, handling emotionally charged queries, and integrating multiple data points for tailored responses. Findings further extend that GenAI interventions can be especially valuable for low SRL students, offering scaffolding that supports engagement with feedback and narrows gaps with their higher SRL peers. At the same time, students' reflections underscored the importance of trust, need for greater adaptivity, context-awareness, and technical refinement in future systems.
Analyzing Message-Code Inconsistency in AI Coding Agent-Authored Pull Requests
Pull request (PR) descriptions generated by AI coding agents are the primary channel for communicating code changes to human reviewers. However, the alignment between these messages and the actual changes remains unexplored, raising concerns about the trustworthiness of AI agents. To fill this gap, we analyzed 23,247 agentic PRs across five agents using PR message-code inconsistency (PR-MCI). We contributed 974 manually annotated PRs, found 406 PRs (1.7%) exhibited high PR-MCI, and identified eight PR-MCI types, revealing that descriptions claiming unimplemented changes was the most common issue (45.4%). Statistical tests confirmed that high-MCI PRs had 51.7% lower acceptance rates (28.3% vs. 80.0%) and took 3.5x longer to merge (55.8 vs. 16.0 hours). Our findings suggest that unreliable PR descriptions undermine trust in AI agents, highlighting the need for PR-MCI verification mechanisms and improved PR generation to enable trustworthy human-AI collaboration.
Excess Description Length of Learning Generalizable Predictors
Understanding whether fine-tuning elicits latent capabilities or teaches new ones is a fundamental question for language model evaluation and safety. We develop a formal information-theoretic framework for quantifying how much predictive structure fine-tuning extracts from the train dataset and writes into a model's parameters. Our central quantity, Excess Description Length (EDL), is defined via prequential coding and measures the gap between the bits required to encode training labels sequentially using an evolving model (trained online) and the residual encoding cost under the final trained model. We establish that EDL is non-negative in expectation, converges to surplus description length in the infinite-data limit, and provides bounds on expected generalization gain. Through a series of toy models, we clarify common confusions about information in learning: why random labels yield EDL near zero, how a single example can eliminate many bits of uncertainty about the underlying rule(s) that describe the data distribution, why structure learned on rare inputs contributes proportionally little to expected generalization, and how format learning creates early transients distinct from capability acquisition. This framework provides rigorous foundations for the empirical observation that capability elicitation and teaching exhibit qualitatively distinct scaling signatures.
From National Curricula to Cultural Awareness: Constructing Open-Ended Culture-Specific Question Answering Dataset
Large language models (LLMs) achieve strong performance on many tasks, but their progress remains uneven across languages and cultures, often reflecting values latent in English-centric training data. To enable practical cultural alignment, we propose a scalable approach that leverages national social studies curricula as a foundation for culture-aware supervision. We introduce CuCu, an automated multi-agent LLM framework that transforms national textbook curricula into open-ended, culture-specific question-answer pairs. Applying CuCu to the Korean national social studies curriculum, we construct KCaQA, comprising 34.1k open-ended QA pairs. Our quantitative and qualitative analyses suggest that KCaQA covers culture-specific topics and produces responses grounded in local sociocultural contexts.
Evaluating Human and Machine Confidence in Phishing Email Detection: A Comparative Study
Identifying deceptive content like phishing emails demands sophisticated cognitive processes that combine pattern recognition, confidence assessment, and contextual analysis. This research examines how human cognition and machine learning models work together to distinguish phishing emails from legitimate ones. We employed three interpretable algorithms Logistic Regression, Decision Trees, and Random Forests training them on both TF-IDF features and semantic embeddings, then compared their predictions against human evaluations that captured confidence ratings and linguistic observations. Our results show that machine learning models provide good accuracy rates, but their confidence levels vary significantly. Human evaluators, on the other hand, use a greater variety of language signs and retain more consistent confidence. We also found that while language proficiency has minimal effect on detection performance, aging does. These findings offer helpful direction for creating transparent AI systems that complement human cognitive functions, ultimately improving human-AI cooperation in challenging content analysis tasks.
Autonomous Agents on Blockchains: Standards, Execution Models, and Trust Boundaries
Advances in large language models have enabled agentic AI systems that can reason, plan, and interact with external tools to execute multi-step workflows, while public blockchains have evolved into a programmable substrate for value transfer, access control, and verifiable state transitions. Their convergence introduces a high-stakes systems challenge: designing standard, interoperable, and secure interfaces that allow agents to observe on-chain state, formulate transaction intents, and authorize execution without exposing users, protocols, or organizations to unacceptable security, governance, or economic risks. This survey systematizes the emerging landscape of agent-blockchain interoperability through a systematic literature review, identifying 317 relevant works from an initial pool of over 3000 records. We contribute a five-part taxonomy of integration patterns spanning read-only analytics, simulation and intent generation, delegated execution, autonomous signing, and multi-agent workflows; a threat model tailored to agent-driven transaction pipelines that captures risks ranging from prompt injection and policy misuse to key compromise, adversarial execution dynamics, and multi-agent collusion; and a comparative capability matrix analyzing more than 20 representative systems across 13 dimensions, including custody models, permissioning, policy enforcement, observability, and recovery. Building on the gaps revealed by this analysis, we outline a research roadmap centered on two interface abstractions: a Transaction Intent Schema for portable and unambiguous goal specification, and a Policy Decision Record for auditable, verifiable policy enforcement across execution environments. We conclude by proposing a reproducible evaluation suite and benchmarks for assessing the safety, reliability, and economic robustness of agent-mediated on-chain execution.
Beyond Static Summarization: Proactive Memory Extraction for LLM Agents
Memory management is vital for LLM agents to handle long-term interaction and personalization. Most research focuses on how to organize and use memory summary, but often overlooks the initial memory extraction stage. In this paper, we argue that existing summary-based methods have two major limitations based on the recurrent processing theory. First, summarization is "ahead-of-time", acting as a blind "feed-forward" process that misses important details because it doesn't know future tasks. Second, extraction is usually "one-off", lacking a feedback loop to verify facts, which leads to the accumulation of information loss. To address these issues, we propose proactive memory extraction (namely ProMem). Unlike static summarization, ProMem treats extraction as an iterative cognitive process. We introduce a recurrent feedback loop where the agent uses self-questioning to actively probe the dialogue history. This mechanism allows the agent to recover missing information and correct errors. Our ProMem significantly improves the completeness of the extracted memory and QA accuracy. It also achieves a superior trade-off between extraction quality and token cost.
Learning Latent Action World Models In The Wild
Agents capable of reasoning and planning in the real world require the ability of predicting the consequences of their actions. While world models possess this capability, they most often require action labels, that can be complex to obtain at scale. This motivates the learning of latent action models, that can learn an action space from videos alone. Our work addresses the problem of learning latent actions world models on in-the-wild videos, expanding the scope of existing works that focus on simple robotics simulations, video games, or manipulation data. While this allows us to capture richer actions, it also introduces challenges stemming from the video diversity, such as environmental noise, or the lack of a common embodiment across videos. To address some of the challenges, we discuss properties that actions should follow as well as relevant architectural choices and evaluations. We find that continuous, but constrained, latent actions are able to capture the complexity of actions from in-the-wild videos, something that the common vector quantization does not. We for example find that changes in the environment coming from agents, such as humans entering the room, can be transferred across videos. This highlights the capability of learning actions that are specific to in-the-wild videos. In the absence of a common embodiment across videos, we are mainly able to learn latent actions that become localized in space, relative to the camera. Nonetheless, we are able to train a controller that maps known actions to latent ones, allowing us to use latent actions as a universal interface and solve planning tasks with our world model with similar performance as action-conditioned baselines. Our analyses and experiments provide a step towards scaling latent action models to the real world.
Safe Continual Reinforcement Learning Methods for Nonstationary Environments. Towards a Survey of the State of the Art
This work provides a state-of-the-art survey of continual safe online reinforcement learning (COSRL) methods. We discuss theoretical aspects, challenges, and open questions in building continual online safe reinforcement learning algorithms. We provide the taxonomy and the details of continual online safe reinforcement learning methods based on the type of safe learning mechanism that takes adaptation to nonstationarity into account. We categorize safety constraints formulation for online reinforcement learning algorithms, and finally, we discuss prospects for creating reliable, safe online learning algorithms. Keywords: safe RL in nonstationary environments, safe continual reinforcement learning under nonstationarity, HM-MDP, NSMDP, POMDP, safe POMDP, constraints for continual learning, safe continual reinforcement learning review, safe continual reinforcement learning survey, safe continual reinforcement learning, safe online learning under distribution shift, safe continual online adaptation, safe reinforcement learning, safe exploration, safe adaptation, constrained Markov decision processes, safe reinforcement learning, partially observable Markov decision process, safe reinforcement learning and hidden Markov decision processes, Safe Online Reinforcement Learning, safe online reinforcement learning, safe online reinforcement learning, safe meta-learning, safe meta-reinforcement learning, safe context-based reinforcement learning, formulating safety constraints for continual learning
Publishing FAIR and Machine-actionable Reviews in Materials Science: The Case for Symbolic Knowledge in Neuro-symbolic Artificial Intelligence
Scientific reviews are central to knowledge integration in materials science, yet their key insights remain locked in narrative text and static PDF tables, limiting reuse by humans and machines alike. This article presents a case study in atomic layer deposition and etching (ALD/E) where we publish review tables as FAIR, machine-actionable comparisons in the Open Research Knowledge Graph (ORKG), turning them into structured, queryable knowledge. Building on this, we contrast symbolic querying over ORKG with large language model-based querying, and argue that a curated symbolic layer should remain the backbone of reliable neurosymbolic AI in materials science, with LLMs serving as complementary, symbolically grounded interfaces rather than standalone sources of truth.
On the Hidden Objective Biases of Group-based Reinforcement Learning
Group-based reinforcement learning methods, like Group Relative Policy Optimization (GRPO), are widely used nowadays to post-train large language models. Despite their empirical success, they exhibit structural mismatches between reward optimization and the underlying training objective. In this paper, we present a theoretical analysis of GRPO style methods by studying them within a unified surrogate formulation. This perspective reveals recurring properties that affect all the methods under analysis: (i) non-uniform group weighting induces systematic gradient biases on shared prefix tokens; (ii) interactions with the AdamW optimizer make training dynamics largely insensitive to reward scaling; and (iii) optimizer momentum can push policy updates beyond the intended clipping region under repeated optimization steps. We believe that these findings highlight fundamental limitations of current approaches and provide principled guidance for the design of future formulations.
Token Maturation: Autoregressive Language Generation via Continuous Token Dynamics
Autoregressive language models are conventionally defined over discrete token sequences, committing to a specific token at every generation step. This early discretization forces uncertainty to be resolved through token-level sampling, often leading to instability, repetition, and sensitivity to decoding heuristics. In this work, we introduce a continuous autoregressive formulation of language generation in which tokens are represented as continuous vectors that \emph{mature} over multiple update steps before being discretized. Rather than sampling tokens, the model evolves continuous token representations through a deterministic dynamical process, committing to a discrete token only when the representation has sufficiently converged. Discrete text is recovered via hard decoding, while uncertainty is maintained and resolved in the continuous space. We show that this maturation process alone is sufficient to produce coherent and diverse text using deterministic decoding (argmax), without reliance on token-level sampling, diffusion-style denoising, or auxiliary stabilization mechanisms. Additional perturbations, such as stochastic dynamics or history smoothing, can be incorporated naturally but are not required for the model to function. To our knowledge, this is the first autoregressive language model that generates text by evolving continuous token representations to convergence prior to discretization, enabling stable generation without token-level sampling.
Differential syntactic and semantic encoding in LLMs
We study how syntactic and semantic information is encoded in inner layer representations of Large Language Models (LLMs), focusing on the very large DeepSeek-V3. We find that, by averaging hidden-representation vectors of sentences sharing syntactic structure or meaning, we obtain vectors that capture a significant proportion of the syntactic and semantic information contained in the representations. In particular, subtracting these syntactic and semantic ``centroids'' from sentence vectors strongly affects their similarity with syntactically and semantically matched sentences, respectively, suggesting that syntax and semantics are, at least partially, linearly encoded. We also find that the cross-layer encoding profiles of syntax and semantics are different, and that the two signals can to some extent be decoupled, suggesting differential encoding of these two types of linguistic information in LLM representations.
Paradoxical noise preference in RNNs
In recurrent neural networks (RNNs) used to model biological neural networks, noise is typically introduced during training to emulate biological variability and regularize learning. The expectation is that removing the noise at test time should preserve or improve performance. Contrary to this intuition, we find that continuous-time recurrent neural networks (CTRNNs) often perform best at a nonzero noise level, specifically, the same level used during training. This noise preference typically arises when noise is injected inside the neural activation function; networks trained with noise injected outside the activation function perform best with zero noise. Through analyses of simple function approximation, maze navigation, and single neuron regulator tasks, we show that the phenomenon stems from noise-induced shifts of fixed points (stationary distributions) in the underlying stochastic dynamics of the RNNs. These fixed point shifts are noise-level dependent and bias the network outputs when the noise is removed, degrading performance. Analytical and numerical results show that the bias arises when neural states operate near activation function nonlinearities, where noise is asymmetrically attenuated, and that performance optimization incentivizes operation near these nonlinearities. Thus, networks can overfit to the stochastic training environment itself rather than just to the input-output data. The phenomenon is distinct from stochastic resonance, wherein nonzero noise enhances signal processing. Our findings reveal that training noise can become an integral part of the computation learned by recurrent networks, with implications for understanding neural population dynamics and for the design of robust artificial RNNs.
WESR: Scaling and Evaluating Word-level Event-Speech Recognition
Speech conveys not only linguistic information but also rich non-verbal vocal events such as laughing and crying. While semantic transcription is well-studied, the precise localization of non-verbal events remains a critical yet under-explored challenge. Current methods suffer from insufficient task definitions with limited category coverage and ambiguous temporal granularity. They also lack standardized evaluation frameworks, hindering the development of downstream applications. To bridge this gap, we first develop a refined taxonomy of 21 vocal events, with a new categorization into discrete (standalone) versus continuous (mixed with speech) types. Based on the refined taxonomy, we introduce WESR-Bench, an expert-annotated evaluation set (900+ utterances) with a novel position-aware protocol that disentangles ASR errors from event detection, enabling precise localization measurement for both discrete and continuous events. We also build a strong baseline by constructing a 1,700+ hour corpus, and train specialized models, surpassing both open-source audio-language models and commercial APIs while preserving ASR quality. We anticipate that WESR will serve as a foundational resource for future research in modeling rich, real-world auditory scenes.
Hybrid Federated Learning for Noise-Robust Training
Federated learning (FL) and federated distillation (FD) are distributed learning paradigms that train UE models with enhanced privacy, each offering different trade-offs between noise robustness and learning speed. To mitigate their respective weaknesses, we propose a hybrid federated learning (HFL) framework in which each user equipment (UE) transmits either gradients or logits, and the base station (BS) selects the per-round weights of FL and FD updates. We derive convergence of HFL framework and introduce two methods to exploit degrees of freedom (DoF) in HFL, which are (i) adaptive UE clustering via Jenks optimization and (ii) adaptive weight selection via a damped Newton method. Numerical results show that HFL achieves superior test accuracy at low SNR when both DoF are exploited.
Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions
We propose Concept Tokens, a lightweight method that adds a new special token to a pretrained LLM and learns only its embedding from multiple natural language definitions of a target concept, where occurrences of the concept are replaced by the new token. The LLM is kept frozen and the embedding is optimized with the standard language-modeling objective. We evaluate Concept Tokens in three settings. First, we study hallucinations in closed-book question answering on HotpotQA and find a directional effect: negating the hallucination token reduces hallucinated answers mainly by increasing abstentions, whereas asserting it increases hallucinations and lowers precision. Second, we induce recasting, a pedagogical feedback strategy for second language teaching, and observe the same directional effect. Moreover, compared to providing the full definitional corpus in-context, concept tokens better preserve compliance with other instructions (e.g., asking follow-up questions). Finally, we include a qualitative study with the Eiffel Tower and a fictional "Austral Tower" to illustrate what information the learned embeddings capture and where their limitations emerge. Overall, Concept Tokens provide a compact control signal learned from definitions that can steer behavior in frozen LLMs.