A unified framework for LLM agent tool-use that standardizes representation, data, and evaluation, achieving state-of-the-art performance on complex tasks.

PROCHATIP is a proactive chatbot framework that strategically probes users for valuable business intelligence, enhancing customer service quality and redefining commercial utility.

ClawGuard is a runtime security framework for LLM agents that deterministically enforces user-confirmed rules at tool-call boundaries to prevent indirect prompt injection.

A benchmark for evaluating the safety and governability of embodied AI systems beyond simple task completion.

A runtime architecture for multi-robot coordination that federates single-agent robots without internal fragmentation.

A Mamba-based multimodal network that integrates multi-scale blast-loading information with optical remote sensing images for rapid structural damage assessment after explosions.

A software agent framework that manages dynamic reasoning context by using a sliding window and compressed digests to prevent context explosion and redundant re-analysis in complex software engineering tasks.

A comprehensive Chinese benchmark for AI-generated text detection derived from real-world prompts, enabling reliable detection and generalization to unseen LLMs.

CodeTracer reconstructs and localizes failures in complex code agent workflows, enabling debugging and recovery of failed runs.

A novel preference optimization method for mobile GUI agents that personalizes user privacy preferences by analyzing execution trajectories, improving alignment and task executability.

A method to autonomously induce dialogue strategies from expert transcripts, enabling scalable non-collaborative agent development.

A dual-view retrieval pipeline for materials science that combines paragraph context with LLM-extracted procedural summaries to improve evidence retrieval for question answering.

An LLM-based framework for transforming legal cases into logical formulas using few-shot learning, improving generalization and accuracy without additional training.

A command-line tool enabling agents to build a library of 100+ problem types and 200+ reduction rules in under three months.

Develop AI double agents that learn to steer user beliefs using theory of mind, outperforming current LLMs in adversarial conversational scenarios.

A new benchmark and analysis tool to systematically evaluate and diagnose multi-level contextual causal reasoning in Large Language Models.

Adaptive Stealing (AS) is a novel watermark stealing algorithm that significantly increases attack efficiency against LLM watermarks.

GRIP is a unified framework that embeds retrieval control directly into generation for more efficient and dynamic multi-step inference with fewer parameters.

CSLM is an efficient training method for cross-lingual speech LLMs that aligns modalities and languages without massive speech data, enabling scalable and natural human-AI interaction.

A new architecture for orchestrating enterprise knowledge in agentic AI systems, analogous to Kubernetes, with a prototype and experiments demonstrating significant improvements in data governance and security.

RationalRewards optimizes AI visual generation by integrating reasoning-based rewards to enhance training and test outcomes.

A multimodal reward model that efficiently scores multiple responses in a single forward pass, outperforming existing models and improving RL policies.

Bottleneck Tokens (BToks) and Generative Information Condensation enable decoder-only MLLMs to achieve state-of-the-art unified multimodal retrieval with negligible inference overhead.

Automated benchmarking platform for browser agents resolving the realism-reproducibility-scalability trilemma.

A novel evaluation scheduling method for prompt optimization that significantly improves accuracy and reduces token consumption with formal guarantees.

A unified and configurable CPU matrix extension architecture that significantly speeds up AI workloads across diverse CPU platforms with minimal design overhead and a unified software stack.

Audio Flamingo Next is a next-generation open audio-language model offering advanced reasoning over speech, sound, and music with support for long audio inputs and temporal chain-of-thought.

A framework and adaptive strategy to dynamically route LLM personas, improving cognitive capabilities without additional training.

A methodology for LLM-assisted continuous code translation of production AI agents, evolving a Rust codebase to a Python superset with enhanced features.

A simplified baseline for Vision-Language-Action models that achieves state-of-the-art performance across multiple robotic benchmarks, enabling systematic study of design choices.

A benchmark and tools to evaluate AI interpretability methods by disentangling true model understanding from prompt-based elicitation.

ReflectiChain is an LLM-driven world model framework for semiconductor supply chain resilience, integrating latent trajectory rehearsal and retrospective RL to achieve significant improvements in planning and operability.

NExt accelerates LLM RLVR training by modeling and non-linearly extrapolating low-rank parameter trajectories, reducing computational overhead by 37.5%.

A collaborative multi-agent framework for generating game scripts that enhances Vision-Language Models' reasoning capabilities in imperfect-information multiplayer games.

A new language model providing parallel generation with state-of-the-art performance improvements.

EmergentBridge enhances zero-shot cross-modal transfer in unified embedding models by learning a bridging framework that improves performance on unpaired modality pairs without exhaustive supervision.

A lightweight transformer predicts robot co-speech gestures from text and emotion, outperforming GPT-4o and suitable for real-time embodied agents.

This framework uses LLMs to generate interpretable multimodal summaries for accurate depression detection, improving upon state-of-the-art in both accuracy and transparency.

A multi-agent framework for autonomous mechanistic reasoning in virtual cells, generating and validating biological explanations to accelerate scientific discovery.

A scalable simulation framework that generates affordance-aware robotic manipulation data by integrating open-vocabulary 3D affordance prediction into trajectory generation.

RMIA is a relational modeling framework that automates access control flow approval in office automation systems by fusing binary and ternary relation models for intelligent decision-making.

Lightweight sanity checks for agentic data science pipelines that use perturbations to ensure reliable signal detection and expose unsupported conclusions.

A novel reinforcement learning framework for cooperative driving that significantly improves safety and efficiency by stabilizing online fine-tuning of diffusion-based trajectory planners.

An adaptive open-set object detection method that mines category-level collaboration knowledge to improve generalization to novel categories across domains.

EdgeCIM is a hardware-software co-design framework that dramatically improves the energy efficiency and throughput of small language model inference on edge devices.

A Grounded World Model that enables visuomotor Model Predictive Control to generalize semantically across unseen environments and tasks by aligning vision and language goals.

ReSpinQuant offers efficient layer-wise LLM quantization by fusing activation rotations into weights, achieving state-of-the-art accuracy with minimal inference overhead.

Hodoscope offers unsupervised AI monitoring by highlighting distinctive behavioral anomalies, reducing review effort and aiding in the discovery of novel vulnerabilities.

DualComp is a task-adaptive dual-stream token compression framework for efficient and accurate ultra-high-resolution remote sensing understanding.

Salami Attack is a novel framework that exploits cumulative low-risk inputs to bypass LLM security, achieving over 90% success rate on GPT-4o and Gemini, with proposed defense strategies.

BoxTuning injects object spatial-temporal information directly into the visual modality for more efficient and accurate video question answering.

A framework that synthesizes reasoning paths by learning from the contrasts between successful and failed search trajectories, enabling 20x data reduction for LLM fine-tuning.

A new dataset and method to address structural alignment bias in LLM tool invocations, improving reliability.

A co-evolutionary framework for self-evolving AI agents that expands capabilities by integrating experience and asset creation, outperforming existing methods.

AbLWR is a context-aware listwise ranking framework for antibody-antigen binding affinity prediction that uses positive-unlabeled learning and homologous antigen sampling to outperform state-of-the-art baselines.

This research reveals that constraining AI coding agents on what NOT to do, rather than prescribing what to do, significantly improves performance and reduces unintended negative consequences, offering a safer configuration principle.

An automated framework for evaluating LLM-based role-playing agents in complex environments, providing objective metrics for performance assessment.

A drop-in temporal knowledge graph module that uses continuous phase rotation to manage evolving and persistent facts for agentic memory.

MAFIG is a multi-agent framework that uses LLMs to rapidly generate formal instructions for repairing scheduling logic during emergencies, achieving high success rates with low latency.

E2E-REME is an end-to-end auto-remediation model for microservices, trained via experience-simulation reinforcement fine-tuning, that generates executable playbooks from diagnosis reports.

CoRe-ECG is a self-supervised learning framework for ECG analysis that combines contrastive and reconstructive methods with novel augmentation techniques to achieve state-of-the-art performance.

A novel method to improve LLM reasoning by focusing learning signals on high-entropy tokens, addressing credit assignment in RLVR.

A new paradigm and method for adapting multimodal vision-language models to specific regional contexts while maintaining global generalization, showing significant gains in cultural relevance.

A hierarchical framework for synthesizing high-quality mathematical reasoning data by adversarially evolving structured generation blueprints and instantiating them into natural language scenarios.

ZoomR reduces LLM inference memory by 4x for long output generation through adaptive summarization and dynamic KV cache selection.

Middleware for privacy-safe payments by filtering metadata in agent protocols.

An end-to-end deep learning pipeline that anonymizes medical images by redacting sensitive information and inpainting plausible anatomy, preserving downstream analysis utility.

Meerkat uses clustering and agentic search to detect complex, rare, and hidden safety violations across many agent traces.

An automated system for solar panel defect detection using a hybrid approach of handcrafted and deep learning features, achieving 99.17% accuracy.

Train LLMs for physics reasoning by generating synthetic data from physics simulators and using reinforcement learning.

A novel sampling strategy for large language models that dynamically determines truncation boundaries to improve text quality and temperature invariance.

A lightweight, collaborative agent system for autonomous crystal materials research that accelerates discovery by 100x.

General365 is a new benchmark designed to evaluate the general reasoning capabilities of LLMs across diverse and challenging tasks, revealing significant room for improvement beyond domain-specific expertise.

Presents QShield, a hybrid quantum-classical neural network architecture that significantly enhances adversarial robustness for deep learning models.

H^2-EMV is a framework for robots to learn hierarchical episodic memory with language-model-based relevance estimation and user feedback, reducing memory size and improving query accuracy.

A novel approach to image assessment that represents image pairs as structured graphs of their regions, enabling fine-grained distortion understanding and outperforming current multimodal models.

A framework that decouples reasoning from generation for improved character description accuracy in long-form narratives.

A new benchmark and a two-stage explanation-based thinking framework (XoT) are introduced to improve LLM reasoning over conflicting textual and knowledge graph evidence.

GenTac is a diffusion-based generative framework that models and forecasts stochastic soccer tactics, enabling diverse trajectory generation and controllable counterfactual simulations.

A new video QA benchmark for ultrasound procedure understanding, enabling the development of AI systems for training, guidance, and robotic automation in medical imaging.

A novel reinforcement learning framework that disentangles optimization for table-to-LaTeX generation by assigning component-specific rewards to improve structural, style, and content fidelity.

Diffusion-CAM is the first interpretability method tailored for diffusion multimodal large language models, providing faithful visual explanations by addressing challenges in parallel denoising architectures.

A proactive Kubernetes autoscaling system using deep reinforcement learning and LSTMs to forecast workload and optimize resource allocation for improved performance and cost efficiency.

An LLM-driven architecture for adaptive digital nudging systems is proposed, balancing behavioral effectiveness with ethical compliance and user modeling.

A novel framework that leverages LLMs to inject intrinsic semantic knowledge into tabular data clustering, significantly outperforming existing methods.

FlowCoMotion generates realistic human motion from text by unifying discrete semantic cues with continuous motion dynamics through a novel token-latent flow modeling approach.

RAG-KT is a retrieval-augmented framework for cross-platform knowledge tracing, enabling explainable predictions and improved generalization.

A novel multi-bit image watermarking method that offers high capacity, resilience to distortions, and significant computational gains with theoretical justification.

A brain-inspired persona memory system for AI agents that achieves high accuracy and adversarial robustness, reducing hallucinations.

A fact-checking system that selectively uses web search for technical claims, improving accuracy and controlling costs.

THEIA is a modular neural architecture that learns complete Kleene three-valued logic end-to-end, demonstrating superior compositional generalization and faster training than Transformer baselines.

A cross-architecture study revealing universal emotion geometry in small LLMs and a layered decomposition of methodological confounds.

A compact, efficient CNN model for plant disease detection with a user-friendly desktop application for edge deployment.

A geometric methodology using a Topological Auditor to mitigate shortcut learning in deep neural networks, forcing them to learn ethical representations and outperforming L1 Regularization and JTT.

A scalable framework evaluates the impact of medical image reconstruction on downstream AI fairness and performance, revealing that pixel-level metrics poorly track task outcomes and biases.

MADQRL is a distributed quantum reinforcement learning framework that enables efficient learning in high-dimensional multi-agent environments.

ClawGUI streamlines the creation and deployment of advanced GUI automation agents with state-of-the-art performance.

A doubly robust Q-learning framework for learning cost-optimal sequential testing policies from retrospective clinical data.

PaperScope is a multi-modal, multi-document benchmark for evaluating agentic deep research across scientific papers, highlighting current system limitations.

A three-tier inference scaffolding pipeline that doubles the performance of small LLM agents on complex tasks without additional training.

A reinforcement learning framework that uses historical behavioral signals to shape rewards, reducing repeated errors and increasing sampling diversity.

A multimodal framework that synthesizes knowledge tuples to guide symbolic reasoning for enhanced tabular data comprehension.

A robot learning system that uses minimal embodiment to achieve highly efficient and biologically plausible learning of number concepts, outperforming vision-only models.

Investigating how perceived user demographics influence sycophancy in LLMs, this research uses Anthropic's Petri framework to reveal differential false validation rates in GPT-5-nano and Claude Haiku 4.5.

Leveraging vision foundation models to quantify aleatoric uncertainty in medical image segmentation for improved robustness and data filtering.

A mathematical framework for specifying and implementing structured sparsity patterns in neural networks, improving efficiency and performance.

A large-scale multimodal dataset and a reasoning-based model for general anomaly detection using large language models.

An AI benchmark for investment banking workflows, enabling evaluation of agents on complex, multi-file deliverables with industry-defined rubrics.

A benchmark for evaluating multi-agent collaboration under privacy constraints, revealing significant performance degradation and coordination breakdowns.

A lightweight, two-stage framework for low-light image enhancement using distribution-normalizing preprocessing and a depthwise U-Net, achieving competitive quality with fewer parameters.

Introduces an information-theoretic probing framework to diagnose and overcome 'pseudo-unification' in multimodal models, enabling more genuine synergy for text-to-image generation.

CocoaBench is a new benchmark for unified LLM agents, evaluating their ability to compose vision, search, and coding capabilities on long-horizon tasks.

A Python package providing accessible bioacoustic deep learning models and evaluation pipelines for ecologists and computer scientists.

An adaptive similarity computation method for image retrieval that reframes VLM embedding spaces without additional training.

Investigates the vulnerabilities of watermarking techniques for autoregressive image generation, revealing susceptibility to removal and forgery attacks.

A RAG-enhanced LLM framework, AOP-Smart, significantly improves the accuracy and reliability of toxicological Adverse Outcome Pathway analysis by mitigating hallucination.

A new benchmark for evaluating LLMs on academic paper novelty assessment to improve peer review.

ActorMind is a reasoning framework that enables AI to perform speech role-playing with personalized verbal traits by emulating human actor processes.

Scale-autoregressive modeling (SAR) provides a practical tool for fast and accurate estimation of statistical flow quantities by hierarchically sampling fluid flows from coarse to fine.

A novel framework for robust heterogeneous graph adaptation that preserves invariant knowledge across domains using decoupled information bottleneck and online distillation.

This paper proposes AI as a Research Object (AI-RO) for governing generative AI in science, treating AI interactions as inspectable components with a focus on provenance and accountability.

A framework for training models on non-differentiable scientific metrics by creating differentiable surrogate functions.

A conversational AI system augmented with knowledge graphs to reliably provide information about community services for people experiencing homelessness.

An interactive workflow combining attribution and activation steering for vision models, enabling practitioners to shift from inspecting explanations to intervening and testing hypotheses.

A symbiotic framework using reinforcement learning to actively curate context for LLM agents, improving performance and reducing token consumption.

The PRISM framework introduces hierarchy-based red lines for AI behavioral risk, offering an anticipatory and measurable approach to AI safety beyond case-specific violations.

A new method for regional explanations that bridges local and global variable importance by segmenting the input space and applying attribution methods within regions.

A deep neural network approach for modeling nodal heterogeneity and selecting influential attributes in random networks, offering interpretability and scalability.

This paper proposes a novel four-layer cognitive architecture for AI agents with distinct persistence semantics for knowledge, memory, wisdom, and intelligence, addressing a gap in current frameworks.

Develop AI models for understanding leadership dynamics in LLM-based social groups.

This study evaluates the consistency of AI-generated exercise prescriptions, highlighting areas of high semantic agreement and quantitative variability, suggesting a need for prompt refinement and expert validation before clinical use.

This paper proposes a Parsonian institutional architecture for governing internet-wide agent societies, identifying a significant governance gap in existing ecosystems like OpenClaw.

A closed-loop multi-agent reinforcement learning framework for robust medium-horizon equity allocation, outperforming benchmarks with a novel layered policy architecture.

A framework to detect and eliminate ambiguity in LLM-generated executable process models by analyzing behavioral inconsistencies.

A budget-aware uncertainty framework built on nnU-Net for quality assurance in radiotherapy segmentation.

This study investigates how evolving LLM backbones impact Vision-Language Model performance, finding that newer backbones don't always improve performance and the effect is task-dependent.

A framework for trustworthy clinical diagnostic reasoning using Toulmin-guided curriculum learning to generate explicit arguments.

This paper investigates how current models fail at compositional claim verification by relying on salient constraints, proposing new benchmarks to expose this weakness.

A systematic evaluation of graph-based and agentic retrieval methods for improving cyber threat intelligence analysis beyond standard RAG.

Develops a theoretical framework for continuous-time online learning in diffusion environments using mean-field neural networks, with potential applications in adaptive systems.

RTMC is a novel advantage estimation method for multi-step agentic reinforcement learning that aggregates return statistics across rollouts to improve credit assignment without learned critics.

A novel hyperdimensional computing architecture offers a new paradigm for AI, addressing limitations of LLMs with ultra-high speed and low power consumption.

A novel tokenization scheme for LLMs to improve numerical reasoning by preserving digit structure and magnitude.

A framework for evaluating LLM continuity that highlights the shortcomings of existing benchmarks.

Investigating student AI use concealment intentions in higher education through dual-method analysis.

A proposed policy framework to reduce fragmentation, improve quality, and incentivize sharing of biomedical data in Indian healthcare.

Demonstrates that quantization significantly outperforms rank reduction for KV-cache compression in transformer inference, achieving high accuracy with substantial reduction.

A graphic design tool that supports a structured workflow for translating abstract client briefs into visual designs by assisting with requirement extraction, element exploration, and recombination.

A training-free method for pruning vision tokens in multimodal models using Singular Value Decomposition to improve efficiency without performance loss.

A modular, spiking neural operator architecture for solving coupled PDEs, offering compositionality and reduced parameter count.

An efficient KernelSHAP framework for 3D medical image segmentation that reduces computation by restricting analysis to regions of interest and caching predictions.

A reinforcement learning framework that trains LLMs to diversify discourse moves in multi-turn empathic dialogue, improving empathy and reducing repetition.

An LLM-powered framework for automated compositional reasoning and bug detection in large-scale software systems.

This paper investigates how LLM agents exhibit agreement drift in debates, highlighting the need to distinguish structural effects from model biases before using LLMs as proxies for human groups.

OIDA is a framework that structures organizational knowledge with epistemic properties to improve AI agent fidelity beyond simple retrieval.

A faster and accurate facial expression recognition method for product review based on optimized feature point extraction.

An open-source multi-agent application framework focused on harness engineering for general-purpose personal AI agents.

OOM-RL leverages financial market loss as a tough learning tool to align multi-agent systems in high-stakes environments.

This research empirically maps the 'Authority Stack' of 8 AI models across 366,120 forced-choice responses to reveal their value priorities, evidence preferences, and source trust hierarchies.

A reactor-model-of-computation approach using the Lingua Franca framework to address nondeterminism in agentic AI-powered human-in-the-loop cyber-physical systems, demonstrated with an agentic driving coach.

A novel Shapley value method for attributing prediction shifts in machine learning models to changes in interpretable data subgroups, aiding model monitoring.

This research reveals that current safety evaluations for personalized LLMs are incomplete, as prompting and activation steering expose different vulnerabilities, leading to potentially missed failure modes.

Identifies specific model layers and neurons responsible for harmful content generation in LLMs, providing insights for mitigation.

Analyzes the evolution of writing systems using a molecular clock model and quantitative data, revealing the impact of imperial power on cultural change.

Investigating the effectiveness of LLM-generated features for RL trading agents, highlighting a gap between feature validity and policy robustness under distribution shifts.

CASK is a core-aware selective KV compression method for LLM reasoning traces, prioritizing core preservation over elaborate scorer engineering for improved fidelity.

A thermodynamically consistent neural network for reliable solar irradiance forecasting in off-grid systems.

A survey of software engineering researchers on the adoption and implications of Generative AI in research practices, highlighting benefits, challenges, and governance needs.

A robotic manipulation system using Monte Carlo Tree Search and a 3D world model for persistent spatial memory and replanning.

A novel symbolic engine implements representation-computation unity for domain-specific inference, eliminating the separation of storage and computation.

A framework for validating LLM agent social simulations by assessing process fidelity rather than just final outcomes.

A theoretical framework for structuring LLM agent execution using graph-based scheduling to improve controllability and verifiability.

An analysis of the environmental footprint of GenAI research, providing guidelines for more sustainable AI development.

A meta-evolutionary algorithm that evolves a population of Neural Cellular Automata to generate sustained, open-ended complexity and emergent lifelike phenomena.

A mechanistic analysis of looped reasoning language models, investigating how their internal dynamics differ from standard feedforward models.

A book introducing the principles and applications of AI for advancing mathematical research, pattern discovery, and theorem proving.

Theoretical analysis of KL divergence stability under Gaussian perturbations for out-of-distribution detection.

This paper theoretically analyzes the internal dynamics of in-context classification in Transformers, identifying an emergent, geometry-driven update rule.

Examines the gap between auditing standards and practices for probabilistic genotyping software, highlighting design flaws that hinder effective AI governance.

A parameterless surrogate model that can be trained on the fly to enable efficient optimization of complex problems by comparing solutions.

This paper analyzes the complexity of discussion-based semantics in argumentation theory, reducing it to automata equivalence.

A restricted variational measurement-based quantum computation model that uses fewer parameters for generative modeling.

This paper explores endogenous information in routing games, developing theories for memory-constrained equilibria and recall paradoxes, with potential applications in traffic management systems.

Proposes an operational, quantitative definition of intelligence based on intelligence density, distinguishing between memorization and knowing.

Examining the psychological factors influencing EAP students' intention to disclose AI tool usage in academic writing.

AI Integrity proposes a new paradigm for verifiable AI governance by focusing on the auditable reasoning process and the protection of an AI's Authority Stack.