PackForcing: Efficient long-video generation using short-video training with reduced memory footprint and improved temporal coherence.

Deploy a lightweight semantic segmentation model for real-time marine environment analysis on resource-constrained devices.

LaMP offers a cutting-edge robotic manipulation framework leveraging 3D scene flow for enhanced vision-language-action alignment, outperforming existing models by integrating geometric foresight in control policies.

AD-CARE: An AI-powered, guideline-driven agent for enhancing Alzheimer's diagnosis accuracy and clinical efficiency through modality-agnostic multi-modal data integration.

LILAC translates natural language into robotic actions using optical flow for efficient task execution.

GridVAD leverages natural-language anomaly proposals for zero-shot video anomaly detection, achieving state-of-the-art performance without task-specific training.

A framework for validating persona agent consistency across multiple dimensions for reliable human simulation.

A novel graph neural network approach for adaptive image compression that significantly outperforms state-of-the-art methods by modeling spatially varying redundancy.

A novel tokenization architecture and algorithm that significantly reduces token count and inference costs for complex scripts, unlocking LLM potential for the Global South.

Leveraging lightweight LLMs as low-latency judges to secure public chatbots against prompt attacks in real-time production environments.

Automatically distill transferable agent skills from execution experience, enabling LLM agents to tackle complex tasks without parameter updates.

ShotStream enables real-time, interactive multi-shot video generation for storytelling by using a novel causal architecture with dual-cache memory and a two-stage distillation strategy.

A feed-forward 3D rendering framework that enables high-fidelity 4K novel view synthesis by predicting compact Gaussian primitives and per-primitive textures, overcoming resolution scaling barriers.

Enhance existing vision foundation models with multi-resolution processing for improved performance across various tasks without retraining.

A framework to improve identity consistency and reduce artifacts in reference-to-video generation by explicitly aligning visual features.

A vision-language-action model that learns to drive by following natural language instructions, enabling personalized autonomous driving experiences.

A personalized vision-language-action driving framework that adapts to individual driver habits and real-time natural language instructions.

An automated graphic design system that emulates human creative workflows to produce production-quality designs from user instructions.

A zero-shot optical flow model leveraging pre-trained vision priors for accurate large displacement estimation and long-range point tracking.

Quantify player skill in table tennis by learning generative models of tactical strokes and embedding them in a latent space that reflects individual characteristics.

Enhance retrieval-augmented generation systems by dynamically distilling and enriching the knowledge base for improved factual accuracy and performance.

A data-driven approach to generate realistic human-object interaction animations by leveraging the denoising process itself for guidance, eliminating the need for explicit classifiers.

A lightweight adapter for multimodal LLMs that enables precise object-centric temporal grounding in videos with improved out-of-domain generalization.

A new benchmark for evaluating and improving AI image generation for commercial design tasks, revealing significant gaps in current models.

A diffusion framework for precise, controllable, and fine-grained facial expression editing with strong identity preservation.

WildASR provides a diagnostic benchmark and analytical tools to improve the reliability and safety of Automatic Speech Recognition systems in real-world voice agents.

A large-scale synthetic dataset and lightweight module to significantly improve 3D hand pose estimation accuracy and generalization for RGB and RGB-D inputs.

An automated data generation pipeline for deformable object manipulation in robotics, enabling scalable robot learning.

Externalize agent harness logic into editable natural language for improved transferability and study, powered by an intelligent runtime.

A new method for vision-language models that improves compositional understanding without sacrificing zero-shot performance, with code available.

Autonomous coding agents that significantly accelerate hardware design optimization by intelligently decomposing and exploring complex configurations.

A novel hybrid memory system for video world models that tracks dynamic subjects even when they are out of view, ensuring motion continuity and realistic simulation.

A training-free framework that uses visual attention to eliminate hallucinations in multimodal large language models by ensuring text generation is visually grounded.

A framework for easily editing object trajectories in videos by specifying a path in a single frame, producing coherent and realistic edits.

Wan-Weaver enables interleaved text and image generation by decoupling planning and visualization, achieving state-of-the-art performance without real interleaved data.

Accelerate LLM generation by training-free self-speculative decoding for diffusion models, achieving significant speedups with improved accuracy.

A framework for autonomous, self-evolving software development using LLM agents and rigorous testing to ensure quality and prevent regressions.

Autoregressive zooming over satellite imagery for precise GPS-denied localization, outperforming existing retrieval methods.

Reinforcement learning for robot world models to enable stable, long-term visual prediction for simulation.

An AI framework using tablet interactions to continuously screen for early cognitive-motor developmental issues in children, identifying persistent deficits for timely intervention.

Enabling customizable speed and overtaking for autonomous vehicles by leveraging existing driving data.

A novel approach to rebalance imbalanced safety data in Cyber-Physical Systems by leveraging behavioral uncertainty, significantly improving safety predictor performance.

Accelerate robot adaptation by decoupling and merging training objectives for enhanced capabilities with reduced computational cost.

Automate the design of complex imaging systems from natural language, overcoming expertise bottlenecks and accelerating scientific prototyping.

A transformer-based metamodel for accurate and efficient 3D atmospheric flow simulation in urban environments, outperforming existing methods.

A framework enabling large multimodal models to perform efficient visual reasoning directly in latent space, improving fine-grained understanding.

A memory-centric social machine learning framework enabling privacy-preserving collaboration among heterogeneous agents by sharing abstracted knowledge instead of model parameters.

A transformer-based AI model that significantly improves radiologist performance in assessing pulmonary nodules, with clinical trial validation.

A masked autoencoder for accurate spatiotemporal forecasting of systems with irregular time steps, applicable to climate and fluid dynamics.

A framework for structured data factorization and iterative learning to significantly improve robotic model generalization with fewer demonstrations.

Real-time neural deformation field for animating complex garment meshes in virtual environments.

A hierarchy-aware multimodal AI that accurately identifies species from imperfect image and DNA data, crucial for conservation and environmental monitoring.

A framework for trustable audio-based depression diagnosis that encrypts sensitive user identity information while maintaining high diagnostic accuracy.

A new benchmark and baseline model for height-aware reasoning in remote sensing, addressing a critical gap in current multimodal models.

A machine learning framework integrating metabolic modeling and optimization to predict and dramatically enhance biofuel-relevant biomass production in yeast.

A multimodal AI system fuses operating microscope and OCT images for real-time, precise instrument tracking and tool-tissue distance estimation in ophthalmic surgery.

A multilingual text-to-speech model that clones voices from just 3 seconds of audio, outperforming existing solutions in naturalness and expressivity.

A framework for scalable, physiologically realistic full-body musculoskeletal motor learning, enabling rapid training and fine-tuning of human-like movements.

Extracts 3D object articulations from unannotated egocentric videos to enable robotics and animation applications.

A self-supervised framework for robust microservice incident management that handles missing data, improving anomaly detection, failure triage, and root cause localization.

A novel training method for video generation models that decouples motion and visual quality, enabling superior performance even with imperfect data.

A new dataset and method for individual bird re-identification and behavior monitoring in the wild, outperforming state-of-the-art.

Lightweight GenAI models generate realistic network traffic for data augmentation and classification, overcoming data scarcity and privacy concerns with high fidelity and efficiency.

A state-of-the-art open-source image restoration model trained on a large-scale dataset, outperforming existing methods and closing the gap with closed-source alternatives.

This research benchmarks and identifies vulnerabilities in LLM-enhanced search engines against sophisticated SEO manipulation, offering insights for building more resilient AI search systems.

A runtime monitoring framework for object detectors that predicts failures by measuring semantic misalignment between internal features and foundation model embeddings, significantly improving recall for safety-critical objects.

EcoThink is an adaptive inference framework that significantly reduces LLM energy consumption for generative AI agents without sacrificing performance, enabling sustainable and accessible AI.

A novel transformer architecture for accurate 3D scene reconstruction from single indoor images, outperforming existing methods.

A model-agnostic framework to extract directed temporal influence structures from pre-trained time series models, improving interpretability and delay localization.

A zero-shot reinforcement learning algorithm that enables direct deployment of policies to real-world robots without finetuning by maximizing exploration entropy.

CIAR accelerates image generation by intelligently offloading computation to the device, reducing cloud requests by 70% and achieving 2.18x speed-up.

A diffusion model that generates safer and more goal-aligned autonomous driving trajectories by decoupling near-term and long-term planning.

A training-free system that uses a second language model to detect confident errors in a primary model, improving LLM safety and reliability.

A globally optimal point cloud registration framework that significantly tightens bounding for faster and more robust results, even with extreme noise and outliers.

A multi-agent reasoning framework and benchmark for explaining diverse micro-video misinformation, enabling robust debunking.

VideoWeaver enables robot policies to transfer to new environments by performing consistent multi-view video translation, overcoming limitations of single-view approaches.

A reinforcement learning framework that uses English as a pivot to significantly improve multilingual mathematical reasoning in LLMs by decoupling understanding from reasoning.

A real-time monitor for LLM reasoning safety that detects and interrupts unsafe reasoning steps, improving security and reliability.

A unified diffusion model for robot manipulation that generates future observations and actions in parallel, improving long-horizon consistency and performance.

Exploit algorithmic side-channels in on-device Vision-Language Models to infer sensitive user contexts, enabling proactive security measures for Edge AI.

A fast Transformer-based segmentation decoder that leverages frozen vision foundation models for efficient image and video segmentation.

A framework for multi-robot coordination that handles dynamic tasks and uncertainties, improving efficiency and reliability.

ALPS automates least-privilege enforcement for serverless functions, reducing security risks and improving permission management across cloud providers.

A lightweight and effective network for infrared small target detection that uses frequency and semantic guidance to improve precision and efficiency.

A phased distillation framework that creates compact synthetic datasets from large image-text data, significantly improving student model performance and reducing storage.

A novel relational distillation framework to create efficient, lightweight CLIP models that preserve structural relationships, outperforming existing methods.

A novel framework for robot navigation that bridges global topological planning with local perception for more efficient and robust object-goal navigation.

A data-driven framework that uses Bayesian learning and deep smoothing to significantly improve the accuracy of inertial-aided navigation systems for robotics and mapping.

A novel framework for flexible and controllable multi-instance sketch video colorization with enhanced detail fidelity.

Develops a novel method for precisely modeling and predicting the difficulty of integer arithmetic puzzles, enabling personalized learning experiences.

An agentic humanoid robot system for autonomous industrial safety, detecting fires, abnormal temperatures, and intruders.

A new benchmark and theoretical framework for evaluating deep research agents, revealing significant gaps in their ability to perform complex structural synthesis.

Leverage off-the-shelf LLMs to compress long texts into compact latent codes, enabling token-efficient long-context reasoning and generation.

A novel method to significantly accelerate 3D vision transformers by intelligently redistributing attention sparsity based on head sensitivity, reducing computational cost without sacrificing accuracy.

A framework that adaptively optimizes document chunking for RAG systems, significantly improving answer correctness and question success rates without changing models or prompts.

A new dataset and benchmark for multi-reference image generation that overcomes current model limitations by providing structured, long-context supervision.

A controllable low-light image enhancement framework with a new dataset and efficient state space model architecture.

A framework and dataset for automatically extracting structured knowledge graphs from scientific papers, enabling deeper document understanding and reasoning.

A multimodal AI framework that constructs 3D channel fingerprints for enhanced 6G communication by fusing diverse data sources, outperforming existing methods by over 27.5% in accuracy.

SliderQuant offers accurate post-training quantization for LLMs by adaptively quantizing layers based on their sensitivity, outperforming existing methods across various tasks and models.

A foundation model for 3D skeletal motion predicts multi-system health phenotypes, enabling gait to be used as a scalable, passive vital sign.

A large-scale real-world dataset and benchmarks for Vehicle-to-UAV cooperative perception, addressing limitations of ground-level autonomous vehicle sensing.

An LLM-in-the-loop framework and dataset to improve hate speech detection by assessing claim check-worthiness, reducing moderator effort and improving model accuracy.

A benchmark for evaluating and improving pragmatic communication and coordination in large language models for complex tasks.

EagleNet enhances text-video retrieval by learning fine-grained relationships between text and video frames, leading to more accurate context-aware embeddings.

ViewSplat enables high-fidelity, real-time 3D scene reconstruction from unposed images by dynamically adapting Gaussian splatting to each viewpoint.

A novel neural compression framework for LiDAR point clouds that achieves real-time speed with competitive compression performance.

A novel vision-based approach for multi-month trajectory anomaly detection that significantly outperforms existing methods and offers substantial speedups.

MolQuest provides a novel agent-based benchmark to evaluate and improve LLMs' abductive reasoning for complex scientific tasks like chemical structure elucidation, revealing significant performance gaps in current frontier models.

A training-free method for out-of-distribution detection that dynamically selects negative labels based on test-time activation to significantly improve accuracy.

FEAST uses a novel attention mechanism and off-grid sampling to infer spatial gene expression from whole slide images, overcoming limitations of existing graph neural networks and providing biologically plausible insights.

Leveraging offline reinforcement learning to outperform classical heuristics on complex combinatorial optimization problems like the Traveling Salesman Problem.

A training-free system for accurate 6D pose estimation of unseen surgical instruments using only CAD models, enabling robust tracking in clinical environments.

A benchmark and baseline framework for evaluating and improving end-to-end automated web testing agents, addressing critical gaps in test completeness and reliability for industrial deployment.

A highly efficient small-target detection framework for anti-UAV surveillance, optimized for edge deployment with a new dataset and improved accuracy.

A foundation model that predicts wireless channel evolution by understanding 3D geometry and signal dynamics, enabling physics-aware 6G intelligence.

A training-free framework that significantly improves long video generation quality from pre-trained short-clip models by adaptively correcting out-of-distribution issues.

An interpretable framework for detecting multimodal misinformation by reasoning over discovered concepts, outperforming existing methods.

A causal framework using conditional instrumental variables to achieve robust domain generalization in medical imaging by disentangling pathological semantics from site-specific artifacts.

A safety alignment technique that reduces harmful LLM outputs in mathematical reasoning without sacrificing accuracy, supported by a new dataset and released code.

A benchmark to evaluate and improve LLM adherence to clinical practice guidelines, crucial for safe healthcare deployment.

A 4D latent diffusion transformer for synthesizing realistic cardiac MRI sequences, improving temporal consistency and physiological accuracy.

AnyID enables ultra-fidelity video generation from any visual reference, offering precise attribute control for creative expression.

Generate privacy-preserving synthetic psychiatric data using LLMs guided by clinical knowledge, overcoming limitations of real data access.

A dual-stage framework that efficiently selects high-utility prompts for reinforcement learning in large language models, reducing computational costs without sacrificing performance.

A novel training framework for machine translation that adaptively leverages document-level context to improve translation quality and robustness.

A controllable 3D medical image synthesis tool using diffusion transformers for improved fidelity and spatial guidance.

A new bilingual text-to-motion benchmark and baseline model that enables high-quality motion generation from diverse language inputs, including zero-shot code-switching.

A novel dual-stream framework for robust egocentric 3D human pose estimation from fisheye cameras, leveraging action-guided motion and kinematic joint encoding.

ET-SAM accelerates scene text detection and layout analysis by efficiently predicting point prompts, enabling faster inference and better data utilization.

A self-supervised learning framework for 3D point clouds that enhances instance localization and semantic understanding, enabling scalable foundation models for diverse downstream tasks.

A dataset and retrieval system for personalized sports skill improvement from instructional videos.

A unified 3D anomaly detection model that disentangles category semantics to achieve state-of-the-art performance and reliability.

A brain-inspired vision foundation model that offers improved data efficiency and interpretability through hierarchical memory mechanisms.

An enhanced RAG framework that improves multi-hop reasoning and domain-specific QA through ontology-guided extraction, advanced clustering, and dual-channel retrieval.

A framework for fine-grained dataset distillation that creates more discriminative and diverse synthetic datasets, improving recognition performance.

A self-supervised framework for detecting audio-visual deepfakes by learning solely on authentic videos, overcoming dataset bias and improving generalization.

A new dataset and code for robust 6D object pose estimation in egocentric views under extreme real-world conditions, addressing limitations of current benchmarks.

A novel Bayesian framework for robust principal component completion that directly identifies sparse component support, improving foreground extraction and anomaly detection.

A framework for robust panoramic semantic segmentation that adapts models from labeled pinhole datasets to unlabeled panoramic data without access to the original source data, improving performance on challenging real-world applications.

Adapt 3D vision foundation models for high-fidelity, pose-free novel view synthesis with a novel alignment and rating framework.

A causal learning framework that improves recommendation systems by intelligently fusing user behaviors and mitigating bias.

A robust and anytime framework for collaborative task sequencing and multi-agent path finding that improves solution quality and success rates in complex environments.

A framework for generating diverse documents from natural language instructions by synthesizing large-scale HTML/CSS datasets and using height-aware reinforcement learning.

A new evaluation framework using signal detection theory to measure LLM metacognitive efficiency, revealing which models truly know what they don't know.

A formula-based data augmentation technique for image classification that procedurally generates Moire interference patterns on-the-fly to significantly improve model robustness with negligible computational cost.

Scales generative multimodal reward modeling using a novel multi-stage reinforcement learning approach, reducing reliance on costly multimodal preference data and significantly improving performance on visual understanding and generation tasks.

A reinforcement learning framework for smarter, modality-balanced data labeling in multimodal AI, reducing annotation costs and improving accuracy.

A framework and dataset for training and evaluating LLMs specifically for mental health conversations, demonstrating superior reasoning and performance.

Leveraging LLMs as adaptive controllers for topology optimization to achieve superior engineering designs.

ElephantBroker provides a verifiable and trustworthy memory system for LLM agents, enabling secure and auditable operation in high-stakes environments.

Pixelis is a pixel-space agent that learns to act and adapt in visual environments by directly manipulating images and videos, enabling grounded multimodal perception and embodied adaptation.

A new benchmark and evaluation framework for multimodal LLMs to detect sophisticated fraud in scientific papers.

A training-free method to reduce hallucination in multimodal LLMs by reinforcing intermediate visual features, improving output reliability without significant computational overhead.

A novel method for training image analysis systems that reliably generalize across different data sources by enforcing feature sparsity and causal interventions.

A plug-and-play strategy to improve multimodal LLM reasoning by dynamically reweighting critical perception and reasoning tokens during training.

A new benchmark and dataset for accurate, machine-readable citation parsing to improve scholarly infrastructure.

This research investigates the link between LLM math problem-solving ability and their accuracy in assessing student reasoning steps, suggesting potential for improved AI tutors.

A cooperative multi-agent reinforcement learning system for fair resource allocation in wireless networks using reconfigurable intelligent surfaces.

A new dataset and multimodal captioning framework for understanding tactical dynamics in full badminton matches.

Augmented reality visualization of impedance control for teleoperation improves force-critical manipulation tasks by providing intuitive, real-time feedback without haptic hardware.

A system design for robots to maintain internal state consistency in long-horizon tabletop games, demonstrated with Mahjong.

A causal inference framework to optimize ICU patient discharge strategies, demonstrated on MIMIC-IV.

A novel low-rank approximation technique for quantized LLMs that reduces overhead and improves accuracy by selectively restoring layers.

A neural framework that directly forecasts probability distributions for dynamical systems, offering improved uncertainty quantification.

A study comparing circular-aware methods for image rotation estimation, identifying the most robust and accurate approaches for vision pipelines.

Leveraging Deep Reinforcement Learning to optimize autonomous vehicle control in mixed traffic, improving road capacity and fuel efficiency.

A novel control method for mobile manipulators that minimizes kinetic energy for safer and more efficient human-robot interaction.

A stateful execution infrastructure for agentic EDA tools to enable iterative optimization and state preservation in production environments.

A new framework for free-form density estimation that estimates the CDF first to overcome limitations of direct PDF estimation, outperforming existing methods.

A new dataset and benchmark for accurately replicating real-world football team tactical behaviors, moving beyond simple reward optimization.

Leverage existing X.509 certificates for privacy-preserving on-chain identity using zero-knowledge proofs.

This paper catalogs and creates new Basque dialectal NLP resources, addressing data scarcity for low-resource languages.

Photon accelerates 3D medical image understanding by efficiently representing volumetric data with adaptive token scheduling for multimodal LLMs.

Enables low-resource language speech evaluation by adapting weakly-supervised ASR models without phoneme time alignment.

A mathematically grounded framework for building multimodal AI systems that can tolerate sensor failures and signal degradation.

A physics-constrained AI framework for rainfall-runoff modeling that improves predictive accuracy and interpretability by embedding hydrological process knowledge.

A reinforcement learning framework that enforces cross-modal consistency to improve multimodal reasoning accuracy.

A universal platform for mobile additive manufacturing robots that integrates navigation and material deposition for adaptable production in dynamic environments.

Develops a framework for understanding and optimizing neural network scaling for weather forecasting, enabling more efficient resource allocation and improved prediction accuracy.

This research evaluates the robustness of LLM-based essay scoring systems to irrelevant factors, suggesting potential for more reliable automated assessment.

This research analyzes shifts in academic writing driven by LLMs, developing a method to detect LLM-generated text and understand their impact on language.

This research investigates how visual versus textual explanations impact user perception in educational recommender systems, offering design guidelines.

A neural implicit method for improved 3D surface and reflectance modeling from sparse and noisy radar data.

This research benchmarks demographic fairness in multimodal LLMs for face verification, revealing significant biases and performance disparities across different models and demographic groups.

This research suggests that simpler, template-based approaches may be more efficient and auditable than LLMs for generating SQL queries in many practical database scenarios.

Improves on-policy distillation for LLMs by addressing empirical failure modes to achieve more stable optimization and better downstream performance.

A new metric to evaluate narrative coherence in vision-language models, revealing systematic differences from human storytelling.

A functional Tucker decomposition method that embeds continuity constraints for improved multidimensional data analysis in classification tasks.

Developing custom vision algorithms to overcome hyperspectral imaging challenges for real-time autonomous driving applications.

This research critically assesses 12 fake news detection approaches, highlighting LLMs' potential for zero/few-shot learning in text-only English scenarios.

Lightweight and interpretable time-series models offer competitive performance for urban air quality forecasting, balancing accuracy and computational efficiency.

A novel data augmentation strategy for low-resource machine translation that aligns with resource gradients, outperforming existing models for specific language varieties.

A novel evaluation method for wildfire prediction models that prioritizes operational decision-making by accurately assessing fire activity and minimizing false alarms.

This paper explores prompt engineering techniques to improve LLM classification accuracy for social science texts, finding that minimal context increases yield the best results.

A framework for structured intent representation in human-AI interaction that improves goal alignment and accessibility across languages and models.

A secure federated RAG system for distributed private data retrieval and confidential LLM inference.

Develops a novel machine learning interatomic potential that accurately predicts material properties by incorporating local curvature information, bridging the gap between theoretical simulations and experimental observations.

A hybrid framework for mobile robots that combines Bayesian inference with deep reinforcement learning to improve autonomous object search efficiency and reliability in partially observable indoor environments.

A server-side control layer for federated learning that uses agentic trust coordination to improve reliability in industrial networks.

A framework for evaluating AI manipulation in context-specific human-AI interactions, revealing domain and geographic differences in manipulative behavior.

This research analyzes how weight pruning in language models affects learned features, revealing that rare features are surprisingly resilient and suggesting pruning acts as implicit feature selection.

A novel physical backdoor attack against deep learning-based radio frequency signal classifiers that bypasses existing defenses.

A novel representation learning approach for constrained motion planning that improves success rates and reduces planning time by intelligently selecting start and goal configurations.

This research validates the effectiveness of Explainable AI metrics by demonstrating that not all functional correctness improvements translate to better human understanding, highlighting a gap for AI explainability tools.

A novel tokenization method for image generation that improves semantic understanding and generation quality.

A novel image sharpening technique that efficiently robustifies deep neural networks against adversarial attacks without requiring complex training or optimization.

A framework for creating more effective adversarial attacks on structured computer vision tasks by aligning input transformations with label transformations.

A publicly available dataset of common skin diseases in Bangladesh, with initial benchmarking, to accelerate AI-powered dermatology diagnostics.

Develops a method to evaluate LLM linguistic generalization using structured natural and synthetic data, highlighting the limitations of synthetic data for real-world understanding.

A novel regression fairness method that enforces distributional parity at specific quantiles or thresholds, offering a tunable trade-off between fairness and accuracy.

Develops intrinsic metrics to contextualize machine translation performance for extremely low-resource languages, addressing variability caused by dataset artifacts rather than model capability.

Accelerate the training of robust machine learning models by safely screening out irrelevant data points.

Enhance infrared object detection robustness by embedding physical thermal radiation knowledge into adversarial training.

Automating LLM prompt optimization for linguistic tasks shows task-dependent results, sometimes matching expert performance but not yet replacing it.

A novel attack method that combines prompt injection and database poisoning to compromise Retrieval-Augmented Generation systems, demonstrating significant improvements over existing methods.

A novel ranking optimization approach for video-text alignment that improves caption generation quality over binary preference methods.

A persistent coverage control algorithm for multi-robot systems to improve solar irradiance prediction accuracy in solar thermal power plants.

A new benchmark for evaluating the reliability of rubric-based instruction following in LLMs, revealing significant performance gaps even in advanced models.

A novel method for few-shot class-incremental learning in tactile sensing that adapts to varying acquisition contexts.

A framework for generating self-evolving LLM agents with formal guarantees of safety and correctness.

Develops a diagnostic framework to understand and control the internal workings of vision-language models by analyzing sparse autoencoder features.

This paper explores transferring knowledge from non-neural machine learning pipelines to neural networks to create unified inference engines for multiple ML tasks.

A novel memory architecture to improve the efficiency and scalability of probabilistic computations for trustworthy AI.

A framework for organizing techniques in self-improving large language models, exploring autonomous data generation, evaluation, and refinement.

A research platform to study how people attribute mental states to robots using LLM-generated explanations.

Optimizing animal back mark design for accurate machine learning-based individual monitoring.

A decidable fragment of Signal Temporal Logic for static verification of safety-critical cyber-physical systems, enabling model checking with SPIN.

A neuroevolutionary approach to design CNNs that are intrinsically robust to adversarial attacks, improving post-attack accuracy without sacrificing clean accuracy.

This research investigates the impact of data scale and class ontology on audio transfer learning performance, identifying key factors for effective pre-training.

A theoretical framework to reduce bias propagation in student-teacher AI models.

This research modernizes navigation for embodied agents to generate semantic scene graphs more efficiently by improving decision-making policies and action formulations.

A novel fuzzing technique for cooperative system and firmware execution to discover vulnerabilities.

This paper theoretically analyzes the regret bounds of a practical Bayesian optimization algorithm, offering theoretical insights into its no-regret properties.

This paper explores a novel physical backdoor attack on deep learning-based automatic modulation classifiers in wireless communications, guided by explainable AI.

This paper explores the usability of passkeys versus passwords in Wi-Fi captive portals, identifying design recommendations to improve user experience and reduce error rates.

This research systematically investigates the impact of participant failures on model quality in cross-silo Federated Learning, providing insights into data skewness and availability patterns.

This paper proposes an adversarial training method to improve the robustness of k-means based resource provisioning in fog networks against evasion attacks.

Formalizing the risks of AI assistance in safety engineering to design shadow-resistant workflows for physical AI systems.

This research identifies a fundamental limitation in current LLMs' ability to maintain consistent internal goals, hindering the development of realistic persona-driven applications.

This paper synthesizes empirical evidence on factors influencing the quality of AI-generated code, highlighting the interplay of human and AI system characteristics.

A theoretical framework for structuring autonomous agent economies with a separation of powers model to address reliability and deception issues.

A theoretical framework for understanding model retraining as approximate Bayesian inference to minimize 'learning debt'.

A theoretical framework for designing AI models that are provably correct and stable before training, eliminating post-hoc verification overhead.

This paper argues that the cognitive benefits of writing are lost when outsourced to AI, proposing new pedagogical approaches and critical literacy skills for the age of generative AI.

A theoretical framework for analyzing the input-output behavior of neural networks by approximating density distributions.

Develops a theoretical algorithm for uniform sampling from complex geometric shapes, with potential applications in statistical inference and machine learning.

A theoretical model to understand how neural networks learn rules and memorize facts simultaneously.

This paper introduces theoretical novel approximation methods for analyzing neural network density distributions, lacking immediate product application.