OneSearch-V2 enhances e-commerce search with reasoning and self-distillation, boosting conversion rates and reducing search biases.

AI tool automating teacher-child interaction quality assessments in Chinese preschools for scalable, continuous monitoring.

LightSplat dramatically speeds up and optimizes 3D scene understanding with a lightweight indexing framework, making real-time open-vocabulary segmentation feasible.

Claudini autonomously discovers advanced adversarial attacks on LLMs, offering cutting-edge cybersecurity solutions.

A world model for autonomous driving that generates physically consistent videos even from challenging or invalid trajectories, improving simulation realism.

PaddleOCR-VL enhances document parsing efficiency by focusing on semantically relevant regions with a coarse-to-fine processing framework.

A multi-agent framework that uses deliberate information asymmetry and reinforcement learning to significantly reduce LLM hallucinations in RAG systems.

LLMpedia empowers enterprises with auditable AI-generated encyclopedic content across diverse topics, enhancing knowledge bases.

VFIG converts rasterized images of complex figures into editable SVGs using a novel vision-language model and a large-scale dataset, bridging the gap for designers and researchers.

A pretraining framework for understanding long-form surgical videos, enabling zero-shot recognition of surgical events and improving multimodal alignment.

Develop an advanced aerial swarm pursuit system using deep reinforcement learning for autonomous navigation in cluttered environments.

OmniWeaving offers a state-of-the-art open-source framework for unified video generation with advanced multimodal composition and reasoning capabilities.

SOMA enhances existing robotic vision-language-action models for robust performance in challenging, out-of-distribution environments without retraining.

A framework and multimodal LLM for objective, scalable, and interpretable assessment of advertising image aesthetics to improve commercial conversion rates.

A self-hosted, specialized text-to-SQL model that drastically cuts API costs and latency for conversational data querying in large-scale applications.

A foundational dataset and unified framework for text-driven quadruped motion generation and control, enabling real-time, expressive robot movements.

A neuro-symbolic reasoning system that significantly enhances multi-modal understanding of tabular data, outperforming existing baselines and rivaling commercial LLMs.

A novel deepfake detection system that leverages intrinsic audio-visual coherence, outperforming state-of-the-art with a new high-fidelity dataset and available code.

A novel AI system for hazard detection in intelligent driving that quantifies risk uncertainty to improve safety and reduce false alarms.

Accelerate diffusion model sampling by orders of magnitude using a novel multilevel approximation method for SDEs.

Accelerate autonomous driving reinforcement learning by 80x using a latent world model that compresses diffusion sampling and enables efficient exploration.

Enhance the reliability of robotic vision-language-action policies by reducing distractors and improving object instance grounding at inference time.

An efficient end-to-end autonomous driving framework that achieves state-of-the-art trajectory planning using spatially-aware and dynamics-informed latent world representations.

Leveraging Vision-Language Models to automate perceptual image quality assessment, outperforming traditional methods in specific attributes and offering a scalable alternative to costly psychophysical experiments.

A geometry-centric framework using graph neural networks to achieve highly accurate 3D reconstruction of deformable surgical tissues, outperforming state-of-the-art and generalizing to unseen domains.

Chameleon provides robots with geometry-grounded episodic memory to overcome perceptual aliasing and improve long-horizon manipulation reliability.

A training-free framework for high-fidelity scene text editing that achieves state-of-the-art results without requiring task-specific data.

A novel defense system that safeguards personal photos from malicious image-to-video generation by operating in both Lab and frequency domains.

A generative pretraining strategy for electronic health records that predicts future patient visits, outperforming existing methods and generalizing to new cohorts.

LensWalk is an agentic framework that allows LLMs to actively control their visual observation of videos, improving understanding accuracy by dynamically planning how they see.

SEGAR enables real-time, temporally coherent augmented reality experiences by selectively generating and correcting future visual frames.

A self-evolving GUI agent that learns from failures to achieve high success rates in mobile automation tasks.

A novel approach to few-shot image classification by mixing and aligning text and image prototypes, outperforming existing methods on benchmarks.

A novel method to distill specialized knowledge from fine-tuned LLMs to new models without access to original training data, using perplexity differences to generate synthetic training examples.

Autonomous agents discover and implement performance-critical AI kernel optimizations, outperforming state-of-the-art expert-engineered solutions.

A novel framework for neural operators that enforces physical constraints for stable and accurate fluid dynamics simulations.

A framework to enhance the Theory of Mind capabilities of multimodal LLMs by aligning visual representations with semantic targets, improving reasoning and explanations in video-based scenarios.

A multi-agent AI system that significantly improves the reliability of medical diagnosis confidence scores, enabling safer deployment in clinical settings.

A novel active learning method for rapidly discovering rare visual categories in large unlabeled datasets with minimal user feedback.

A specialized AI model for agentic software engineering that demonstrates strong long-term planning and coding intelligence, with benchmark results comparable to state-of-the-art systems.

A multimodal AI system that reunites lost pets with their families using both visual and acoustic biometrics, inspired by animal communication.

A novel agent system that uses a sorry-driven formal decomposition strategy to efficiently solve complex mathematical reasoning problems in automated theorem proving.

A novel deepfake video detection framework that leverages rich vision-language semantics and identity-aware prompting to significantly outperform state-of-the-art methods.

A large-scale dataset and benchmark for training and evaluating computer-use agents that automate complex desktop workflows.

Leveraging transfer learning on a new gait dataset to create a scalable, non-invasive frailty assessment tool for aging medicine.

A highly efficient generative model for accurate global weather forecasting up to 30 days, outperforming larger models with faster inference.

This paper proposes and experimentally validates a quantum-resistant security mechanism for 5G non-3GPP access by integrating Quantum Key Distribution with IPsec, offering faster and more secure connections.

ClawKeeper provides comprehensive, real-time security for autonomous agents by integrating skill, plugin, and watcher-based protection mechanisms.

A real-time framework for optimally assigning and generating collision-free trajectories for large drone swarms, enabling complex aerial light show choreography.

An autonomous AI research supervision framework that uses a persistent knowledge graph and multi-agent system to discover, develop, and validate AI research.

A plug-and-play module that significantly enhances the 3D spatial understanding of Vision-Language-Action models for improved robotic control.

Generate realistic 3D human-object interactions by extracting visual priors from 2D images using a diffusion-based framework.

An LLM-guided evolutionary search framework that autonomously discovers interpretable molecular optimizations by planning chemical symbolic operations.

A dynamic routing framework for image geolocalization that adaptively selects the best paradigm (retrieval or generation) for precise GPS coordinate prediction.

A highly efficient 5M-parameter OCR system that rivals billion-parameter models by focusing on data quality and diversity, offering superior localization and reduced hallucinations.

A decentralized AI framework for optimizing city-wide traffic signals using novel state representation and neighbor-aware policy optimization.

A novel framework that uses LLMs to distill knowledge into simpler RL agents for more efficient and generalizable traffic signal control.

A language-guided network that uses text prompts to improve the detection of camouflaged objects in images.

A novel approach inspired by brain plasticity to significantly improve the efficiency and accuracy of deepfake audio detection models.

A new benchmark and dataset for evaluating multimodal LLMs in 3D virtual agents, revealing significant gaps in current AI capabilities for embodied perception and reasoning.

A multi-modal self-supervised learning framework that unifies RGB and LiDAR depth representations for improved performance on downstream driving tasks.

An AI-powered scheduler that autonomously optimizes the learning order of semantic classes for improved image segmentation under challenging conditions.

A neighborhood-adaptive linear regression method to refine low-resolution time-space traffic diagrams, improving accuracy and capturing complex traffic dynamics.

A new parallel dataset for Hindi-Sanskrit translation, enabling significant performance gains for contemporary language models.

A novel state space model that refines forgotten specifics for state-of-the-art video semantic segmentation with high computational efficiency.

VERIA enhances 3D object detection for rare classes in autonomous driving by synthesizing diverse and contextually relevant multimodal instances using off-the-shelf foundation models.

TopoMesh enables high-fidelity 3D mesh generation by unifying topological representations between ground truth and predicted meshes, leading to superior preservation of sharp features and geometric details.

A novel framework for image clustering that leverages language to create more discriminative features and adaptive semantic centers, outperforming state-of-the-art methods.

ScrollScape enables ultra-high-resolution (32K) image generation at extreme aspect ratios by leveraging video diffusion priors for structural integrity.

A dual-branch Transformer fusion framework predicts anticancer drug response from multi-omics data, outperforming baselines and providing biological explanations.

A plug-in method that enhances time series forecasting accuracy by aligning intermediate representations with pretrained foundation models.

A novel body-part-aware Mixture-of-Experts framework for highly accurate micro-action recognition, outperforming state-of-the-art on challenging benchmarks.

A super-resolution framework that preserves fine-grained details and semantic consistency at the instance level for real-world images.

A novel framework to defend machine learning models against adversarial attacks using GAN-generated synthetic data, improving robustness without sacrificing performance.

UniScale co-designs data and model architecture to unlock superior performance in search ranking by synergistically scaling both elements.

An adaptive, auditable vision-language model for medical diagnostics that generates and executes code for iterative reasoning.

An AI-powered multi-agent system for automated penetration testing of robotic systems, offering high reliability and traceability.

This research quantifies and addresses fairness disparities in Retrieval-Augmented Generation (RAG) systems, offering a path to more equitable LLM applications.

Develops novel attacks to uncover privacy vulnerabilities in time series imputation models, enabling more secure AI deployments.

A dual-sided framework for stable personalized federated learning that enhances client specificity and global model accuracy through hierarchical alignment and adversarial knowledge transfer.

Enhance knowledge distillation by using dual-modality teachers with visual prior enhancement and text-guided adaptive fusion to significantly improve student model performance.

A multi-resolution Transformer architecture that improves time-series forecasting accuracy and robustness by integrating both point-wise and patch-wise temporal representations.

A novel reward modeling framework for super-resolution that aligns with human perceptual preferences by using the low-resolution input as a semantic anchor, supported by a new large-scale dataset and code release.

Enable on-chip training of weightless neural networks for efficient edge AI with TsetlinWiSARD, offering significant hardware efficiency improvements.

Unlock state-of-the-art few-shot and zero-shot image classification for Large Vision Language Models by intelligently combining their internal representations.

Walma uses machine learning to detect memory corruption and tampering in WebAssembly applications, offering a practical solution for runtime integrity verification.

A multi-agent framework for automating complex, long-horizon computer tasks in healthcare, outperforming existing models.

Develops a closed-loop simulation for trajectory prediction that trains autonomous agents to react and recover from their own errors, significantly improving collision avoidance.

A novel neural operator that significantly accelerates PDE solving by decoupling linear and nonlinear effects, offering faster training and comparable or better accuracy.

A reinforcement learning framework that dynamically optimizes deepfake detection training by prioritizing challenging samples for improved generalization.

A frequency-selective filtering framework that amplifies action-specific frequencies to improve skeleton-based action segmentation accuracy and boundary definition.

A novel motion planning method that significantly reduces trajectory computation time for autonomous robots by decoupling safety constraints, enabling faster and safer navigation in complex environments.

A multilingual, multi-turn medical dialogue dataset and model for accessible preliminary healthcare consultations, leveraging parameter-efficient fine-tuning for broad deployment.

A novel graph convolutional network integrating reservoir computing for improved graph classification and generation, with faster convergence and reduced over-smoothing.

A unified framework to analyze and mitigate gender bias in LLMs by correlating internal representations with expressed outputs, demonstrating that current debiasing methods don't fully remove latent bias.

This research identifies and quantifies an 'alignment tax' causing response homogenization in LLMs, proposing a novel uncertainty estimation method to improve selective prediction accuracy and reduce computational costs.

A 2.5D framework for accurate retinal layer segmentation in OCT images, improving glaucoma assessment by fusing cross-slice features for better consistency and robustness.

KCLNet provides an electrically-aware graph representation learning framework for analog circuits, enabling significant performance improvements in critical EDA tasks like classification and subcircuit detection.

A physics-informed neural network that leverages Lagrangian dynamics for highly accurate and discriminative skeleton-based action segmentation.

A training-free method to control lighting in text-to-image generation by manipulating initial noise, offering dynamic user guidance.

A multi-task reinforcement learning framework for robotic manipulation that leverages real-time 3D scene graphs and knowledge to achieve robust generalization and sample efficiency.

This research identifies and quantifies new safety risks in multimodal large language models for image generation, showing they are more prone to generating unsafe content and harder to detect than diffusion models.

A benchmark and diagnostic tool for building generative AI that understands and creates visually compelling posters with human-centered design principles.

A multimodal AI framework for Alzheimer's diagnosis that integrates neuroimaging and clinical data with explicit guideline adherence, offering transparent and accurate diagnostic reasoning.

A lightweight decoding-time intervention method to significantly reduce object hallucinations in Large Vision-Language Models, improving their reliability for critical applications.

A novel transformer architecture that mitigates optimization collapse in visual forensics, achieving state-of-the-art generalization for deepfake detection.

A novel hierarchical graph-enhanced model for improved map-matching performance using self-supervised and supervised learning.

A framework for generating realistic financial tool-use dialogue data to improve LLM capabilities in complex financial scenarios.

A reinforcement learning framework for humanoid robots that allows real-time adaptation of control objectives based on user preferences, enabling diverse and optimized behaviors.

A novel framework for hyperspectral image classification that dynamically routes spatial-spectral features through expert networks to overcome limitations of existing methods.

A training-free method for diffusion model style transfer that preserves content identity and captures complex styles, outperforming existing approaches.

A self-interpretable neural network framework that identifies minimal representations to improve prediction accuracy and uncover human-interpretable patterns.

SemLayer reconstructs editable semantic layers from flattened vector icons, enabling advanced design workflows.

SpectralSplats offers robust 3D object tracking by using frequency domain supervision to overcome vanishing gradients in differentiable rendering, enabling recovery from severe misalignments.

This research introduces a novel training framework to significantly improve the robustness of speech-to-LLM models against noisy and error-prone contextual information during inference, leading to more reliable real-world performance.

A generative framework using Lagrangian relaxation and SDEs to produce diverse, high-quality solution candidates for mixed-integer linear programming, outperforming existing ML baselines and achieving competitive optimality with exact solvers at reduced computational cost.

An embodied agent framework that continuously learns from environment interaction to improve task execution and generalize to new tasks.

A novel open-vocabulary multi-object tracking system that learns from continuous video data and synergistically combines detection and association for improved performance on unseen objects.

A multi-agent LLM framework that personalizes urban sensing tasks for participants, improving satisfaction and fairness.

A RAG-powered AI system for accurate and reliable Islamic inheritance calculations and heir allocation, outperforming state-of-the-art.

A new large-scale dataset and parameter-efficient framework for underwater video object segmentation that significantly outperforms existing methods.

A personalized adaptive clipping framework for federated learning that significantly improves accuracy and convergence speed while maintaining differential privacy.

A novel zeroth-order optimization framework for training high-dimensional Physics-Informed Neural Networks with significantly reduced memory and computational complexity.

A novel training framework that dynamically allocates computational depth in Transformers to reduce training FLOPs by up to 19% while improving performance.

A feed-forward architecture that reconstructs 3D hand meshes and camera poses from uncalibrated images, outperforming state-of-the-art and generalizing to real-world scenarios.

Enabling real-time, safe humanoid robot teleoperation through advanced inverse kinematics and visual pose estimation.

An ensemble of specialized LLMs with a rule-based orchestrator and interpretable student model for reliable, controllable, and efficient adaptive tutoring.

A RAG framework that reconstructs multi-source web documents at a concept level to improve factual consistency and answer quality in Q&A systems.

A real-time action detection system that distills motion knowledge into RGB models, achieving state-of-the-art performance on a single CPU.

Leverage LLMs to create portable patient embeddings from clinical time series data, enabling scalable deployment of predictive models across hospitals with reduced retraining.

A training-free method to efficiently prune LLMs by merging task-specific importance scores, improving accuracy without retraining.

A real-time text-driven humanoid control system that generates physically feasible and safe motion trajectories by integrating physics-guided generation with a multi-stage safety gate.

Leveraging Large Language Models to improve pedestrian gait recognition by encoding binary silhouettes into a language-like representation.

A novel architecture for collaborative perception that robustly handles agent heterogeneity without retraining, enabling scalable and cost-effective multi-agent systems.

Predict material properties of 3D assets from a single image 120x faster than existing methods, enabling real-time physics simulation and digital twins.

A retrieval-augmented generation framework that grounds Arabic LLMs in historical lexicographic data to significantly improve understanding of religious texts.

A tool that predicts and optimizes LLM inference costs by analyzing token consumption, addressing the 'price reversal phenomenon' where cheaper listed models can cost more.

An LLM-powered framework integrated with Splunk to automate and prioritize threat hunting for overwhelmed SOC analysts.

An open-source simulation platform for developing and testing autonomous vehicle control systems using monocular vision.

A novel reinforcement fine-tuning method for 3D point cloud models that significantly improves performance in data-scarce scenarios.

A large synthetic benchmark and baseline methods for more practical multi-view crowd counting and localization, advancing research towards real-world applications.

Automate the discovery of improved policy optimization algorithms for language models, significantly boosting performance on complex reasoning tasks.

A text-to-text generation approach for argument mining that simplifies the process and achieves state-of-the-art results.

A novel deep learning method for audio fingerprinting that handles variable-length audio, outperforming existing state-of-the-art on real-world datasets.

A benchmark and analysis tool to evaluate and improve the context-aware speech generation capabilities of omni-modal AI models.

An AI assistant that generates targeted, evidence-based questions during physician-patient encounters to improve the implementation of medical guidelines.

A novel decoding technique that significantly reduces multi-view hallucination in vision-language models by suppressing visual interference.

A framework to protect private photos from being exploited by vision-language models through data poisoning, preventing identity-affiliation leakage.

A training-free framework that improves occlusion accuracy in text-to-image generation by arbitrating object attention, enhancing compositional capabilities of existing diffusion models.

A multimodal deception detection system that provides auditable reports and robust cross-cultural generalization, powered by novel reasoning datasets and advanced representation learning.

Optimize large vision-language models for faster and more memory-efficient inference, enabling longer contexts and higher throughput.

Accelerate LLM inference by up to 220% with a self-distillation method that improves multi-token prediction accuracy and efficiency.

An AI-powered multilingual interface that automatically enhances text with contextually relevant pictograms to support reading comprehension for children with special educational needs.

A framework for building and deploying ophthalmology-specific multimodal large language models that outperform existing general and medical models.

A Markovian framework to audit the reliability and oversight costs of agentic AI in enterprise workflows.

Optimizing document chunking for Retrieval-Augmented Generation in specialized enterprise domains like oil and gas to improve information retrieval accuracy and reduce computational costs.

This research uncovers how steering LLM behavior can inadvertently increase vulnerability to jailbreak attacks, highlighting a critical trade-off between controllability and safety.

A new method to analyze and measure adaptive scaffolding in tutoring dialogues by aligning semantic content, applicable to both human and AI tutors.

A novel sequential neural policy for safer and more efficient non-linear model predictive control, reducing computational burden and improving feasibility.

An uncertainty-aware transfer learning framework for accurate and reliable lithium-ion battery state-of-health forecasting.

A data-driven surrogate using LSTMs to predict and analyze critical ship roll instabilities from wave data.

A decoding approach for Sanskrit poetry generation that improves semantic coherence and metrical adherence by segmenting verses and using phonetically aware transliteration.

A framework for analyzing the co-occurrence of sentiment, personality, and topics in online discourse to understand audience reactions to virtual influencers.

Develops novel methods to measure and improve fairness in recommender systems by analyzing recommendation parity directly, rather than relying on representation-level proxies.

A survey proposing Causal Transfer Learning to build more robust and generalizable AI models for medical image analysis, addressing domain shift issues.

Develops a specialized reward model for AI math tutors to improve pedagogical quality and mistake remediation.

A novel musculoskeletal robotic hand design that uses electrohydraulic actuators for inherent compliance and self-sensing, eliminating the need for external sensors for safe manipulation.

Automated reward design for cooperative multi-agent systems using LLMs to improve coordination and task performance.

A novel reservoir computing approach simplifies audio signal processing for real-time, energy-efficient speech analysis in embedded systems.

A training-free framework to accelerate diffusion-based video editing by intelligently caching context tokens, reducing redundant computations without sacrificing quality.

A probabilistic time series forecasting framework that explicitly models time-varying variance for more robust predictions and uncertainty quantification.

A framework to integrate sociolinguistic variation into NLP models, improving robustness and performance on evolving languages.

A remote attestation protocol to detect Rowhammer attacks using commodity hardware features and TPMs.

A framework that injects heuristic reasoning priors to improve data efficiency in referring object detection for low-data environments.

A plug-and-play algorithm for Poisson inverse problems that leverages pre-trained denoisers for improved performance in medical imaging applications.

A framework for efficiently tuning control policies by combining numerical data with human preferences using multi-modal surrogate models.

A language-level safety constraint for probabilistic programming languages that prevents models from generating invalid programs, demonstrated with a modification to Stan.

A novel framework for more stable and accurate crowd counting across different environments by discovering and leveraging latent domains.

A causal inference framework to disentangle shared and private information in multiplex graphs for improved representation learning.

A new benchmark and method for diagnosing specific conceptual misunderstandings in students, going beyond simple correctness prediction in educational AI.

An iterative unsupervised framework that jointly performs feature selection and clustering for high-dimensional complex data, outperforming classical and deep baselines.

A dual-branch Swin Transformer network for more accurate road extraction in optical remote sensing imagery.

Enables multiple drones to cooperatively follow complex 3D paths in low-altitude airspace, overcoming disturbances and obstacles with real-time replanning.

This research provides a data-driven taxonomy and analysis of reinforcement learning environments, identifying a paradigm shift towards LLM-driven agents and domain-specific generalization to guide the design of next-generation simulators.

A knowledge-enhanced machine learning model for accurate deep-sea cold seep stage inference using microbial data, reducing reliance on costly visual surveys.

A robot system that proactively assists human manipulation by inferring goals from observed state changes, rather than waiting for explicit instructions.

Generate realistic and varied NPC daily activity plans for immersive digital environments using a novel generative model.

A framework to evaluate LLM judges for code applications by identifying systematic biases and misalignment with human preferences.

This research investigates the limitations of RAG systems in policy analysis, finding that component improvements don't guarantee better answers.

Developing robust multimodal speaker identification systems that perform well even with missing visual data and across different languages.

We propose a system to make autonomous agent memories tradable assets by binding them to verifiable computational provenance and creating a market layer for their exchange.

A Vision Transformer model leverages Sentinel-2 time series and spatial context to discriminate between organic and conventional farming systems, with varying accuracy across crop types.

This research analyzes bias in Automatic Speech Recognition systems for Newcastle English, revealing socially patterned errors and advocating for dialect-aware development.

A new framework for evaluating AI uncertainty attribution methods to enable more reliable and comparable XAI development.

Develops a miniature, fiber-reinforced soft bending actuator for endoluminal robotic interventions, demonstrating significant bending angles and structural integrity.

Develops design principles for causal machine learning-powered clinical decision support systems to enhance trust and collaboration.

A mixture-of-experts architecture for causal discovery from observational data, aiming to overcome limitations of existing methods.

A novel curriculum learning loss function for large-scale speaker verification that significantly improves accuracy by adaptively ranking sample difficulty.

A probabilistic framework models continuous-time evolution of glucose distributions for improved diabetes monitoring and treatment analysis.

A teacher-student diffusion model for generating realistic 3D hand motions from text, improving VR and robotics applications.

A novel algorithm for federated classification that simultaneously ensures privacy and fairness, with theoretical guarantees and experimental validation.

This research explores a novel reinforcement learning approach for translating natural language mathematics into formal proof language, demonstrating improved performance over supervised methods.

A neuro-symbolic system integrates physiological signals for interpretable human fatigue detection, offering insights into model reasoning.

Developing generalist neural motion planners to overcome limitations in cluttered robotic manipulation environments.

A novel approach to improve Graph Neural Network generalization on out-of-distribution data by integrating causal representation learning and loss replacement.

A large-scale study and dataset of Telegram bots to identify and combat illicit activities, providing insights for content moderators and researchers.

A medical image fusion model that embeds copyright identifiers and requires authentication for high-quality outputs, protecting intellectual property.

A GNN layer that intelligently routes messages based on edge type to improve performance on specific types of heterophilous graphs.

A tool to detect structural indicators of security risks in software dependencies.

A novel multitask learning approach for semantic alignment across historical language stages, providing a baseline and guidance for modeling under data constraints.

A system for resolving inconsistent software mentions across scientific documents using semantic embeddings and density-based clustering.

A real-time compiler for dynamic AI models that significantly speeds up compilation and improves model efficiency.

A no-code tool for analyzing citation networks to understand the origin of scholarly influence.

A novel black-box attack method that generates stealthy injection payloads to compromise LLM agents by treating payload generation as a tree-structured search problem.

A novel synthetic data generation pipeline for Reinforcement Learning to improve large language models for code and math tasks.

Enabling on-chip learning for neuromorphic systems with a feedback-control optimizer.

A standardized benchmark suite for multi-objective search to enable robust and reproducible cross-study comparisons.

Optimize sensor placement for graph neural networks to significantly improve leakage detection in water distribution networks.

A bio-inspired hierarchical reasoning framework for LLMs that uses strategic dormancy and mnemonic encoding to improve complex, multi-domain problem-solving.

A routing-guided framework for LoRA fine-tuning of Mixture-of-Experts models that significantly reduces trainable parameters and training time by focusing on the most activated experts.

A framework for open-vocabulary temporal action detection that uses LLM reasoning to decompose actions into phases for better generalization to unseen actions.

This research provides forensic investigators with the tools and methodologies to uncover digital evidence from local LLM runners, addressing a critical blind spot in digital investigations.

A novel neural network architecture inspired by Kirchhoff's laws for improved data representation and physical consistency in tasks like PDE solving and image classification.

A watermarking framework for copyright protection, manipulation localization, and high-fidelity face content recovery to combat deepfakes.

This research analyzes the consistency of Google Play's Data Safety disclosures against actual app behavior, revealing significant inconsistencies in privacy reporting that necessitate improved validation mechanisms.

This paper theoretically analyzes and experimentally validates improvements to minimax search algorithms for two-player perfect information games, enhancing their ability to find winning strategies.

A theoretical Bayesian optimization method for high-dimensional constrained problems.

A novel metaheuristic inspired by free-market economics for open-ended complex system optimization.

This paper investigates why self-distillation can harm LLM reasoning, particularly in mathematical tasks, by suppressing uncertainty expression.

A novel neural network architecture for contextual regression that improves efficiency and interpretability by separating context identification from context-specific regression.

A theoretical framework for improving local intrinsic dimensionality estimation using bagging to reduce variance and mean squared error.

This research explores a theoretical framework for identifying emergent 'self' concepts in continually learning robots by analyzing invariant subnetworks.

This research explores the coordination challenges faced by multiple LLM-based code agents when implementing shared code components, highlighting the critical role of detailed specifications in achieving integration accuracy.

This paper explores optimizing multilingual LLMs using federated learning by analyzing the impact of client language composition on model performance and fairness.

This research identifies a fundamental flaw in how stance detection is currently performed, showing that existing methods fail when attitudes are complex and multi-dimensional.

A new NMF framework that makes basis vectors as distinct as possible to improve interpretability and handle highly mixed data.

Analyzing the geometric properties of word embeddings to understand antonym relationships.

A theoretical framework for improving visual-inertial navigation through equivariant filter transformations.

Enhancing the explainability of image geolocalization models by combining activation maps from multiple network layers.

A lightweight, ML-based security framework for Industrial IoT that predicts and mitigates network condition impacts on trust convergence.

A new framework for LLM training that aims to improve reasoning by better utilizing and internalizing experience during reinforcement learning.

This paper identifies key challenges in setting up iterative generative optimization loops with LLMs, offering guidance for practitioners but not presenting a ready-to-deploy solution.

A quantum-classical generative adversarial network for seismic data processing that leverages quantum mechanics to overcome the representational limitations of traditional neural networks.

A novel framework integrates wireless communication with multi-robot task assignment and route scheduling for enhanced efficiency in smart factories.

Develops theoretical convergence guarantees for gossip algorithms applied to pairwise machine learning objectives in distributed sensor networks.

Develops theoretical optimal regret bounds for infinite-horizon reinforcement learning problems.

A theoretical overview of conformal prediction methods for statistical inference.

A formal framework for understanding self-reference in language, offering a structural basis for semantic knowledge.

This paper theoretically investigates uniform convergence phenomena in product spaces, extending Vapnik--Chervonenkis theory with a focus on linear VC dimension.