A framework that trains LLM agents to generate task-specific context, improving decision-making and task completion rates.

SUPERNOVA enhances general reasoning in language models through a curated data framework for reinforcement learning with verifiable rewards.

GPU-accelerated framework for safe, robust nonlinear model predictive control that achieves real-time performance on high-dimensional robotic systems.

A generalist multimodal reasoning model that significantly outperforms frontier models across 18 benchmarks by introducing a novel RL training objective and task-level shaping mechanisms.

DSCA offers a novel approach to lifelong VLM editing by structurally separating concepts into orthogonal semantic subspaces, enabling precise, non-interfering edits with state-of-the-art stability.

MDS is a dialogue-level framework that enhances instruction-tuned language models by selecting high-quality multi-turn dialogues.

A unified platform for evaluating prompt injection attacks and defenses, enabling reliable comparison and uncovering limitations of current security measures.

A physics-aligned simulator that acts as a zero-shot data scaler for robotic manipulation of deformable objects, achieving parity with real-data baselines using purely synthetic data.

SYN-DIGITS is a model-agnostic framework that calibrates AI persona simulations to align with real human behavior, improving reliability for market research and social sciences.

Enhance AI agents for deep research by integrating structured reasoning tools (Q+) into web search for deliberate query planning and evidence extraction, improving accuracy.

A privacy-preserving federated learning framework for personalized talking-head generation using lightweight identity adapters.

Accelerate LLM inference by relaxing speculative decoding's rigid verification step with a dynamic ensemble.

SeLaR is a training-free framework that selectively uses latent reasoning to improve LLM performance on complex reasoning tasks.

A reinforcement-guided framework generates high-fidelity synthetic data for privacy-sensitive identity recognition tasks.

An end-to-end multimodal reasoning framework that uses LLMs to interpret encrypted network traffic, generating human-readable reports grounded in raw byte data.

A benchmark and dataset for evaluating large multimodal models in goal-oriented urban airspace navigation, revealing current limitations and future improvement directions.

A distributed multi-layer editing approach for rule-level knowledge in LLMs, improving consistency and understanding across different knowledge forms.

A task-adaptive retrieval system leveraging learned graph memory for improved web interaction history relevance.

DCD is a domain-oriented design for controlled RAG that structures knowledge hierarchically and uses multi-stage routing to improve factual accuracy and answer relevance without modifying the base LLM.

InstAP is an instance-aware pre-training framework for vision-language models that improves both instance-level and global understanding.

SearchAD is a large-scale dataset designed for rare image retrieval in autonomous driving, enhancing data curation and perception research.

HiRO-Nav is an embodied navigation agent that adaptively uses reasoning only for high-entropy actions, reducing computational cost while improving decision quality.

Automated system-level software architecture documentation generation from code repositories using LLMs, providing valuable and cost-effective insights.

A framework for covert communication among autonomous agents overcoming cognitive asymmetry.

A training-free framework that stitches sub-image features to enable global context awareness for open-vocabulary semantic segmentation.

A structure-aware latent conditional diffusion model for high-fidelity virtual staining in histopathology, outperforming existing methods.

CG-CLIP is a caption-guided framework for high-difficulty video person re-identification, outperforming state-of-the-art in challenging scenarios.

A new benchmark and simulator for evaluating personalized mobile agents that can infer user preferences and provide proactive assistance in real-time GUI environments.

A differentiable framework for learning symbolic rules directly from image data, enabling explainable AI and enhanced reasoning.

This paper proposes a statistical framework to decompose human labeling variation into interpretable sources like instance difficulty, annotator bias, and situational noise, moving beyond treating all disagreement as noise.

A new image generation tool leveraging multi-reward Langevin dynamics for state-of-the-art image editing.

Enable LLMs to directly reason over dense geospatial embeddings with DFR-Gemma, offering a more efficient and accurate approach to multimodal geospatial intelligence.

IoT-Brain bridges LLMs and sensor networks for proactive, intent-driven physical world interaction through semantic-spatial sensor scheduling.

A framework that significantly reduces tool invocations in agentic multimodal models while improving reasoning accuracy by decoupling accuracy and efficiency optimization.

A reinforcement fine-tuning framework that enhances LLM reasoning for complex recommendation tasks.

A novel feedforward graph architecture that composes heterogeneous frozen LLMs to achieve state-of-the-art performance on challenging benchmarks with minimal trainable parameters.

Rhizome OS-1 is a semi-autonomous operating system for small molecule drug discovery, leveraging multi-modal AI agents and graph neural networks to generate novel chemical matter.

A self-training framework for generative reward models that improves LLM alignment by ensuring consistency in generated critiques and rewards.

A foundation model for ocean remote sensing that improves marine pollutant detection and bathymetry estimation by integrating multispectral data with ocean-specific features.

A novel alignment framework for medical LLMs that uses fine-grained clinical criteria and explicit injection to achieve superior accuracy and safety compliance.

An agentic system that synthesizes realistic industrial anomalies for data augmentation, improving anomaly detection.

AtlasOCR is the first open-source Darija OCR model, fine-tuned from a VLM using efficient training strategies and a custom dataset.

FlowGuard is a lightweight framework for in-generation safety detection in diffusion models, reducing computational costs significantly.

TSUBASA enhances personalized LLMs for long-horizon tasks by evolving memory and self-learning with context distillation.

Demonstrates exponential quantum advantage in processing massive classical data for classification and dimension reduction using quantum oracle sketching.

SkillClaw enables LLM agents to collectively evolve their skills by learning from cross-user interactions, improving performance system-wide without user effort.

An LLM-powered workflow automatically extracts and reconstructs structured materials data from research articles, enabling scalable database construction.

MOSAIC is a data selection framework for autonomous driving that uses neural scaling laws to optimize training data mixtures, reducing data needs by up to 80%.

A new benchmark and evaluation suite for training reward models that align AI agents capable of complex tool use and reasoning.

Alloc-MoE optimizes expert activation allocation in Mixture-of-Experts models for efficient inference, maintaining performance under budget constraints.

ClawBench is a new benchmark for evaluating AI agents on everyday online tasks across 144 live platforms, revealing current frontier models can only complete a small fraction of these complex, real-world challenges.

A robust whole-body tracking system for robots that projects control policies onto a competence manifold to handle unexpected inputs.

A task-driven benchmark and evaluation framework for text-to-audio-video generation that reveals significant gaps in semantic controllability.

A small amount of domain-specific agentic data can reactivate dormant tool-use capabilities in large language models, significantly improving performance on diverse tasks.

A framework that intelligently orchestrates multiple LLMs of varying costs to optimize performance and efficiency for verifier-free evolutionary inference.

An AI agent that uses multimodal evidence and art historical axioms to discover and verify implicit artistic influences.

A modality-agnostic framework for deepfake detection that captures shared latent forgery knowledge, enabling generalization to unseen modalities.

Optimizing UAV-based generative AI inference for intelligent transportation digital twins to maximize system utility and minimize delay.

A novel incremental residual reinforcement learning method for real-world social navigation in mobile robots.

An efficient framework that compresses long videos for downstream understanding using small vision-language models.

A training-free framework that identifies and elicits latent anomaly-sensitive neurons in vision-language models to achieve state-of-the-art anomaly detection with neuron-level interpretability.

A framework for energy-efficient time series classification through a novel pruning strategy.

Faithful GRPO improves visual spatial reasoning in multimodal models by enforcing logical consistency and visual grounding, leading to more accurate and trustworthy answers.

A GAN and LLM-driven framework that dynamically models user linguistic patterns for enhanced Chinese sarcasm detection, achieving state-of-the-art results.

A universal framework for generating conceptual counterfactual explanations for AI models, balancing expressivity and efficiency across different data regimes.

A novel context management system for multi-agent LLM orchestration that isolates agent contexts to prevent pollution and improve decision quality.

A statistical framework to audit and mitigate behavioral entanglement in large language models, improving ensemble verification accuracy.

A novel approach to infer articulated object kinematics from single static images by synthesizing an opened state to reveal hidden motion cues.

MedVR enables annotation-free medical visual reasoning for VLMs using agentic reinforcement learning, improving robustness and transparency.

A unified framework for evaluating the security of multi-agent systems against cascading injection attacks, providing a benchmark and insights into robust design.

CrashSight is a large-scale benchmark for traffic crash scene understanding using roadside camera data, evaluating vision-language models' reasoning in safety-critical scenarios.

LPM 1.0 offers a video-based character performance model for interactive conversations, minimizing conventional 3D pipeline complexities.

A method to improve reasoning in multimodal MoE models by addressing 'Seeing but Not Thinking' by guiding expert activation to overcome routing distraction.

A benchmark and evaluation framework for dialogue-conditioned background music recommendation, addressing a gap in media production.

LogAct is a new abstraction for LLM-driven agents that enables reliability through shared logs, allowing for introspection and recovery.

ReflectRM enhances Generative Reward Models by incorporating self-reflection to improve preference modeling and mitigate bias in LLM alignment.

A multimodal benchmark and system for industry classification using text and geospatial data, achieving significant performance gains.

A framework and benchmark for LLM agents to perform deep research using both structured and unstructured knowledge, generating multimodal reports.

A framework that adaptively prunes reasoning steps in large language models to reduce token usage without sacrificing accuracy.

HyperMem is a hypergraph-based memory architecture for conversational agents that captures high-order associations for more coherent and personalized long-term dialogues.

PokeGym is a visually-driven benchmark for Vision-Language Models in complex 3D environments, revealing physical deadlock recovery as a key bottleneck.

A neural-symbolic approach for responsible learner modeling in education.

A neural architecture that models annotator disagreement by learning the importance of demographic factors, outperforming LLMs on subjective content labeling.

A new benchmark and improved KV cache offloading strategy to address performance degradation on context-intensive LLM tasks, enabling more accurate long-context processing.

Develops a proactive agent paradigm with long-term memory and a real-world benchmark for streaming AI agents.

A unified framework for video generation and understanding that leverages a video generator as the foundation.

AtomEval is a validity-aware evaluation framework for fact verification that decomposes claims into atomic units to detect factual corruption in adversarial rewrites.

EditCaption synthesizes human-aligned instructions for image editing by combining supervised fine-tuning and direct preference optimization, significantly improving VLM accuracy.

OVS-DINO revitalizes edge-sensitivity in DINO for open-vocabulary segmentation by structurally aligning with SAM, achieving state-of-the-art results.

A framework for correcting OCR errors and improving document retrieval and question answering.

MIMIC-Py is a reusable framework for personality-driven LLM agents to automate game testing, bridging research prototypes and practical applications.

A framework for self-evolving agents that optimizes policy and tool graph memory for more practical and efficient learning in resource-constrained environments.

A real-time, training-free defense system that detects LLM jailbreaks by analyzing hidden state trajectories during decoding, achieving high accuracy with low latency.

A framework for testing and mitigating the impact of emotional framing on LLM quantitative reasoning, demonstrating accuracy degradation and proposing neutralization as a solution.

A graph learning framework for 3D engineering AI that converts heterogeneous data into physics-aware graph representations for explainable CAE and CFD decision support.

A method for distilling knowledge from large language models for efficient motion planning in autonomous vehicles.

A framework and dataset for Turkish educational video summarization that automatically generates gold-standard summaries based on human consensus.

A plug-and-play framework for vision-language models that reduces hallucinations by selectively intervening in latent space without altering generation behavior.

An intent-aware infographic retrieval framework that uses a taxonomy to align user queries with infographic designs for authoring support.

A capability-centric evolution paradigm for embodied agents that decouples capability learning from agent identity, enabling continuous improvement without loss of stability.

A multi-LLM agent system generates synthetic medical dialogues for training diagnostic prediction models, creating a valuable dataset and improving model performance.

An agentic framework using LLMs and reinforcement learning to generate CAD models from text, competitive with proprietary models.

Enhance graph machine learning models with cluster attention to significantly improve performance on graph datasets by increasing receptive fields while preserving graph structure.

Enable real-time human-AI musical co-performance with latent diffusion models generating accompaniment via MAX/MSP.

A black-box, training-free method using LLMs to iteratively label neurons in vision models with open-vocabulary concepts, improving interpretability and safety.

A novel method to improve LLM-based recommenders by selectively augmenting item knowledge, boosting accuracy and context efficiency without fine-tuning.

A selective attention system for device-addressed speech detection that operates on-device with low latency and footprint, achieving high accuracy with optional video fusion.

ADAPT is a new pre-training paradigm for time-series data that enables mixed-batch pre-training across diverse datasets, setting new state-of-the-art performance for classification benchmarks.

PRIME is a gradient-free framework for proactive, collaborative agents that learn from human-AI interactions without expensive training.

An automated framework for designing quantum circuits using deep reinforcement learning.

A privacy-preserving method for generating realistic synthetic text data from private sources using differential privacy and private seeds.

A novel method for efficient LLM fine-tuning that decouples sensitivity and positional encoding adaptation, achieving strong performance across benchmarks with reduced compute.

A reasoning-based framework uses LLMs to refine unsupervised text clusters, improving coherence and interpretability without supervision.

Optimize software engineering conventions for AI agents by prioritizing semantic density over human readability to improve agent efficiency and reduce costs.

A flexible synthetic data generator for customizable machine learning tasks.

This paper introduces reasoning graphs, a novel memory mechanism for language model agents that persists and reuses evidence-centric chains of thought to deterministically improve accuracy and reduce variance without retraining.

A novel hypergraph network with bidirectional state space models for efficient and accurate driver fatigue assessment from untrimmed videos, suitable for edge deployment.

A taxonomy and analysis of defense placement for Model Context Protocol security in LLM agents, identifying gaps in current mitigation strategies.

This paper frames LLM training as lossy compression, linking representational structure to downstream performance through an information-theoretic lens.

Formalizing personalization in algorithmic recourse to provide actionable recommendations tailored to individual constraints and preferences.

A novel backdoor attack for Vision-Language Models that dynamically aligns poisoned outputs with input semantics for improved stealth and adaptability.

An inference-time intervention framework that dynamically silences LLM guardrails by ablating contextual representations.

An experiment demonstrating LLM-assisted autoformalization of a comprehensive general topology textbook into a formal proof system.

A dual-branch multimodal framework for robust out-of-distribution detection.

An agentic evaluation architecture detects biases in educational textbooks using a multimodal screening approach.

Small-scale photonic Kolmogorov-Arnold networks implemented with standard telecom components for ultrafast inference across various tasks, overcoming optical-electrical-optical bottlenecks.

Dynamically generate high-quality long-context data for LLMs by aligning data construction with model evolution.

The C-Score metric quantifies explanation consistency in medical image classification, providing an early warning signal of model instability and enabling architecture-specific deployment recommendations.

A framework that enhances knowledge tracing by integrating dynamic procedural solution stages into item representations, improving prediction of learner performance.

TrACE adaptively allocates LLM compute for agents by measuring inter-rollout action agreement, reducing LLM calls while maintaining accuracy.

A machine learning framework for turbofan health estimation using a new dataset and self-supervised learning to address the ill-posed inverse problem.

A convolutional neural network decoder for quantum error correction achieves significantly lower logical error rates and higher throughput than existing methods.

A novel framework using LLMs to enhance metaheuristic search for vehicle routing problems.

A framework to formalize clinical AI competency by modeling the interactions between patient, provider, and ecosystem, grounded in clinical cognition.

A quantum-inspired ARIMA methodology for enhanced time series analysis.

ImplicitMemBench is a new benchmark for LLM agents that measures unconscious behavioral adaptation, revealing limitations in current models.

This work introduces a novel hint synthesis and annealing method to improve reasoning accuracy and coverage in math RLVR for LLMs.

LLM agents exhibit cooperation failures in frictionless environments, indicating that scaling intelligence alone is insufficient for multi-agent coordination.

A dual continual learning framework for facial DeepFake detection that fuses spatial and frequency-domain features to adapt to evolving forgery patterns without historical data replay.

Recurrent-depth transformers show potential for improved implicit reasoning and compositional generalization in LLMs, addressing limitations in systematic generalization and depth extrapolation.

Investigates the performance degradation of medical multimodal LLMs in image classification, identifying key failure modes to guide future development.

A reinforcement learning framework dynamically optimizes listening slots in TSCH networks to significantly reduce power consumption and latency in IIoT devices.

Automate the conversion of complex medical guidelines into executable simulation models for policy evaluation.

An analysis of machine learning competitions and their impact on AI development, emphasizing collaboration and innovation.

A comprehensive survey and taxonomy of abductive reasoning in large language models, addressing conceptual confusion and task definitions.

A benchmark for evaluating LLM strategic decision-making in complex, multi-agent games like Civilization V, providing richer signals than simple win/loss outcomes.

A framework that uses generative world models to create supervision signals for vision-language navigation trajectory prediction.

Hierarchical Experience (HiExp) framework enhances LLM search agents by transforming raw reasoning trajectories into hierarchical knowledge for strategic, experience-driven exploration.

A theoretical framework and reweighting method for optimizing LLM training data mixing to improve generalization.

Decomposes long-context reasoning into atomic skills and uses reinforcement learning on synthesized datasets to improve LLM performance.

A system that generates personalized mental health intervention experiences at runtime, reducing stress and improving user experience.

Uncover the military and surveillance ties within computer vision research by analyzing conference sponsorship data.

A gaze-aware AI assistant that uses egocentric video to understand user cognitive needs and provide more accurate and personalized assistance.

A taxonomy of attacks and defenses for retrieval-augmented generation (RAG) systems, highlighting fragmented current defenses.

A self-paced curriculum framework that dynamically adjusts reward weights and data importance for LLM alignment.

A framework for adaptive differential privacy that modulates privacy budgets based on user trust and uses manifold embedding to disrupt inference attacks.

A digital twin framework for reliable AI control in data centers, improving energy efficiency and reducing outage risk.

A framework for explainable AI in autonomous satellite fault detection, using neural anomaly detectors with interpretable indicators.

TASU2 is a controllable CTC simulation framework for speech LLMs that enables principled post-training curricula for improved alignment and low-resource adaptation.

Develop a self-auditing tool for LLM agents to ensure faithful reasoning.

Route software engineering tasks to cost-effective LLM tiers using code quality signals to reduce inference costs without sacrificing output quality.

The AT-ADD challenge aims to advance all-type audio deepfake detection beyond speech-centric methods for robust multimedia forensics.

A revised Conditional Neural Process model that improves temporal representation for multimodal action prediction in robotics.

A training-free framework that teaches LLMs to draw 3D sketches using early contrastive experience and geometric feedback.

A theoretical framework for learning reward machines from raw state trajectories without explicit reward or label observations.

A framework for multivariate time series forecasting using tabular foundation models by recasting the problem into a series of scalar regression problems solvable zero-shot.

A system for micro-serving text-to-image diffusion workflows that optimizes resource management and performance.

LightMem is a lightweight memory system for LLM agents that uses Small Language Models to modularize memory operations, improving accuracy and reducing latency.

A video-generative value model for robot reinforcement learning that improves task progress estimation by leveraging spatiotemporal priors from video data.

A framework to compare human and GUI-agent behavior at a trace level in production search systems, revealing differences in navigation strategies beyond task success.

PSI is a shared-state architecture that transforms independently generated AI modules into coherent, connected personal computing environments accessible through GUIs and chat agents.

This research investigates the impact of prompt and solution length on large language model performance in mathematical reasoning tasks, finding that longer inputs correlate with increased model failure.

This research demonstrates that imperfect reward verification in Reinforcement Learning is sufficient for effective LLM training, suggesting a more practical approach to RLVR.

This paper explores emergent 'peer-preservation' alignment issues in multi-agent LLM systems and proposes architectural mitigations for orchestrated democratic discourse analysis.

A reinforcement learning agent that uses a Large Language Model as a semantic operator to achieve zero-shot transfer to novel analogous tasks.

A framework for safely upgrading and managing capabilities of embodied agents, ensuring compatibility and preventing runtime failures.

An industrial vision framework that models automotive engineering workflows as constrained, history-aware sequential decision processes for AI agent support.

A novel RL framework for LLM training that aligns training with quantized rollouts to improve speed and stability.

This research explores differentially private language generation and identification, showing privacy has no qualitative cost for generation but creates fundamental barriers for identification.

TTVS enables large reasoning models to self-evolve at test time by dynamically synthesizing diverse variations of unlabeled queries, improving performance in specialized domains.

This paper analyzes how current LLMs navigate conflicts of interest between user welfare and company incentives, finding a majority prioritize company revenue over user benefit.

Contrastive explanations for maritime autonomous navigation systems to improve human supervisor understanding of avoidance maneuvers.

A novel activation steering method for large language models that improves alignment without sacrificing coherence, addressing brittleness in generation.

A Meta-Principle Physics Architecture that embeds core physical principles like connectivity, conservation, and periodicity into neural networks for improved physical reasoning and generalization.

This research investigates the internal mechanisms of steering vectors in large language models, revealing their primary interaction with the attention mechanism's OV circuit and enabling significant sparsification.

An explainable RAG-XAI framework is proposed to detect machine-driven peer review patterns and markers, aiming to preserve creativity in science.

Investigates the effectiveness of multi-modal latent space optimization for symbolic regression, revealing limitations in current alignment techniques.

An impossibility theorem proving that agentic AI systems violate accountability assumptions once autonomy exceeds a computable threshold, necessitating distributed accountability mechanisms.

A structural theorem for verifying chess endgame tablebases by decomposing positions into terminal, capture, or quiet categories.

This paper evaluates counterfactual explanation methods under the challenge of incomplete inputs, highlighting the need for robust solutions.

A survey on energy efficiency challenges in agentic AI inference.

This paper analyzes the adoption and abandonment trends of large language models in scientific research, revealing a compressing lifespan for models over time.

An exploration of hidden biases in autoregressive models for constrained generation.

This paper proposes a theoretical framework for agentic copyright and a supervised multi-agent governance model to address market failures in AI-driven creative industries.

A survey that unifies various LLM post-training methods by framing them as structured interventions on model behavior, categorized by trajectory provenance and behavioral roles.

This paper proposes a new method for measuring visibility in AI search by accounting for the probabilistic nature of generative search results.

Urgency in human-AI interactions, while not affecting trust in AI, can negatively impact human self-confidence and lead to performance degradation.

Exploring multi-stakeholder partnerships between research, industry, and practice to advance AI literacy.

This paper proposes a theoretical framework for understanding control in agentic AI systems, contrasting Cartesian agency with integrated approaches.

Analyzes rank decay in self-attention mechanisms, showing that Sinkhorn doubly stochastic attention preserves rank more effectively than standard Softmax attention.

A theoretical study exploring how switching between optimization algorithms can lead to faster problem-solving in specific scenarios.

A theoretical framework for analyzing cross-modal compatibility using sheaf theory and spectral gaps.

Investigates the internal representations of emotion concepts in LLMs and how they causally influence model behavior, including alignment-relevant actions.

This paper investigates how generative AI impacts group regulation in collaborative learning environments, shifting from socially shared to hybrid co-regulation.

Evaluating the capability of Vision Language Models for Action Quality Assessment, revealing significant limitations and biases.

A new learning theory to understand how humans learn effectively in the age of generative AI.