CodeTracer reconstructs and localizes failures in complex code agent workflows by parsing run artifacts into a traceable state history.

Build a universal problem-solver reducer library for NP-hard challenges, scalable through AI agent integration.

TorchUMM is a unified codebase for evaluating, analyzing, and post-training diverse unified multimodal models across various tasks and datasets.

A glyph-driven fine-tuning framework (GEVO) to enhance multimodal LLMs for ancient Chinese character evolution analysis, releasing a benchmark and trained models.

GRIP is a unified framework that embeds retrieval control directly into generation, enabling end-to-end coordination for question answering.

A unified and configurable CPU matrix extension architecture that significantly boosts AI workload performance across diverse architectures with minimal overhead.

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

An LLM-based framework that transforms legal cases into logical formulas using few-shot learning, improving generalization and interpretability in legal reasoning.

Secure AI agent payments with PII-safe x402 metadata filtering.

This research reveals significant vulnerabilities in existing watermarking techniques for autoregressive image generation, showing they are susceptible to removal and forgery attacks, impacting content attribution and dataset filtering.

A co-evolutionary framework that integrates experience memory and asset memory to enable self-evolving AI agents with expanded capabilities and guided asset creation.

A unified framework for LLM agent tool-use that standardizes data, representation, and evaluation, significantly outperforming commercial models.

An executable benchmark for evaluating knowledge graph readiness in gap and overlap analysis for policy documents.

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.

PROCHATIP is a proactive chatbot framework that intelligently probes users for valuable business intelligence, significantly outperforming reactive support tools in both information gathering and service quality.

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, moving beyond simple task completion metrics.

A runtime architecture for multi-robot coordination that maintains individual robot autonomy, improving governance and recovery containment.

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

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 Grounded World Model that enables visuomotor Model Predictive Control to generalize to unseen environments and instructions by aligning vision and language.

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

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

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

Develop AI agents that learn to steer user beliefs by understanding and manipulating their intentions, outperforming current frontier models in adversarial scenarios.

A benchmark and tools to evaluate AI interpretability methods by breaking the elicitation confounder, showing when white-box methods truly add value.

METER is a new benchmark and analysis framework for evaluating multi-level contextual causal reasoning in LLMs, with code and dataset available.

An adaptive watermark stealing algorithm for LLMs that dynamically selects attack perspectives to significantly increase steal efficiency against target watermarks.

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

A new architecture for orchestrating enterprise knowledge in agentic AI systems, akin to Kubernetes for containers, with proven security and freshness benefits.

A large-scale multimodal dataset and baseline model for general anomaly detection using MLLMs.

Hodoscope offers unsupervised AI monitoring by detecting behavioral anomalies, reducing review effort and discovering novel vulnerabilities.

A benchmark for evaluating AI agents in end-to-end investment banking workflows, designed with input from 502 investment bankers.

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

A novel dataset and method to address structural alignment bias in LLMs' tool invocation, improving reliability.

Introspective Diffusion Language Models improve efficiency and quality of language model generation by ensuring introspective consistency.

RationalRewards teaches reward models to provide multi-dimensional critiques, improving visual generation at both training and test time through reasoning.

An open-source framework simulating LLM social groups with elected leadership, demonstrating significant improvements in cooperation and social welfare for common-pool resource management.

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

Lightweight sanity checks based on the PCS framework to ensure agentic data science pipelines reliably distinguish signal from noise, exposing unsupported conclusions and improving trustworthiness.

An intelligent evaluation scheduling system for prompt optimization that significantly improves accuracy and reduces token consumption.

A novel framework that models and extrapolates low-rank parameter trajectories non-linearly to accelerate LLM reinforcement learning with verifiable rewards.

A lightweight, two-stage framework for low-light image enhancement that achieves competitive quality with fewer parameters.

An automatic framework for multi-turn LLM jailbreaking that chains low-risk inputs to bypass security and achieve high attack success rates.

A cross-language code analysis framework that grounds LLM reasoning in execution-based validation to ensure trustworthy vulnerability detection and repair.

Introduces an information-theoretic framework to diagnose 'pseudo-unification' in multimodal models, revealing divergence in encoding and response patterns to enable genuine synergy.

Audio Flamingo Next is an open-source audio-language model capable of understanding and reasoning over 30 minutes of speech, sound, and music, with a novel temporal reasoning paradigm.

A framework for synthesizing high-quality mathematical reasoning data by adversarially evolving constraint graphs, outperforming existing datasets.

A Python package that makes bioacoustic deep learning models accessible for analysis and research.

Leveraging physics simulators and reinforcement learning to train LLMs for deep physical reasoning, achieving significant zero-shot sim-to-real transfer on physics benchmarks.

A novel sampling strategy for LLMs that dynamically adjusts truncation based on logit distribution to improve text quality and temperature invariance.

A methodology for LLM-assisted code translation and evolution of production AI agents, demonstrating code reduction and feature expansion.

An interactive workflow combining activation steering with XAI for instance-level analysis in vision models, enabling practitioners to shift from inspection to intervention-based hypothesis testing.

A framework for Tool-Internalized Reasoning (TInR) that integrates tool knowledge directly into LLMs, improving reasoning efficiency and performance without external documentation.

ReflectiChain is an LLM-driven world model framework that enhances supply chain resilience by integrating latent trajectory rehearsal and retrospective agentic RL for autonomous policy evolution.

A collaborative multi-agent framework for generating game scripts to enhance vision-language models' reasoning in imperfect-information multiplayer games.

A framework evaluates how medical image reconstruction impacts downstream AI diagnostic performance and fairness, revealing that pixel-level metrics don't capture task performance or bias amplification.

EmergentBridge improves zero-shot cross-modal transfer in multimodal embedding models by learning a bridging framework that strengthens connections between unpaired modalities without requiring exhaustive pairwise supervision.

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

A multi-stage framework using LLMs for interpretable multimodal depression detection generates clinical summaries to guide fusion of text, audio, and video features, improving accuracy and transparency.

A multi-agent framework for autonomous mechanistic reasoning in virtual cells, generating and validating biological explanations to improve 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 modeling for intelligent decision-making.

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

Bottleneck Tokens (BToks) and Generative Information Condensation enable unified multimodal retrieval by providing explicit pooling and token-level supervision for semantic compression.

An adaptive open-set object detection system that mines category-level collaboration knowledge to generalize to novel categories with significant performance gains.

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

ClawGUI is a cutting-edge framework for building and deploying efficient GUI automation agents.

ReSpinQuant offers efficient layer-wise LLM quantization by fusing activation rotations offline and matching bases in residual subspaces, achieving state-of-the-art accuracy with minimal overhead.

DualComp is a task-adaptive dual-stream token compression framework for ultra-high-resolution remote sensing understanding, improving efficiency and accuracy.

BoxTuning injects object spatial-temporal information directly into the visual modality for multimodal models, reducing token costs and improving video question answering accuracy.

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 actions, significantly improves performance and reduces reliability risks, offering a clear principle for safer agent configuration.

An automated framework for evaluating LLM-based role-playing agents, providing objective metrics for complex, constraint-heavy environments.

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

PaperScope is a multi-modal, multi-document benchmark for evaluating agentic deep research systems across scientific papers.

A test-time adaptation framework to bridge the open-loop to closed-loop gap in autonomous driving policies.

MAFIG is a multi-agent framework that uses LLMs to rapidly generate formal instructions for repairing scheduling logic during emergencies, reducing 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 framework and optimization technique to improve LLM reasoning by addressing the credit assignment problem in RLVR through entropy analysis.

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

A novel approach to reduce LLM inference memory by adaptively compressing reasoning thoughts and dynamically selecting KV cache details, achieving over 4x memory reduction.

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

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

A framework and adaptive strategy to induce and dynamically route personas in LLMs, showing stable shifts in cognitive capabilities and outperforming static personas.

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

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

A lightweight collaborative agent system for crystal materials research that significantly outperforms larger models and accelerates discovery.

A robot learning system that uses minimal embodiment to achieve highly efficient acquisition of abstract number concepts, mirroring human cognitive development.

General365 is a new benchmark designed to rigorously assess and improve the general reasoning capabilities of large language models across diverse, challenging tasks.

Presents QShield, a hybrid quantum-classical neural network architecture that enhances adversarial robustness of deep learning models by integrating quantum circuits for feature encoding.

A novel approach to image assessment that uses distortion graphs to represent region-level degradation, outperforming current multimodal LLMs and offering a new direction for fine-grained image analysis.

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

A new benchmark and a two-stage reasoning framework address LLMs' struggles with conflicting knowledge from text and knowledge graphs, improving faithful reasoning.

A conversational AI system that uses knowledge graphs to provide reliable, location-aware information about community services for people experiencing homelessness.

GenTac is a diffusion-based generative framework that models and forecasts soccer tactics as a stochastic process of player trajectories and tactical events.

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

A novel reinforcement learning framework that disentangles optimization across table structure, style, and content to improve fidelity in table-to-LaTeX generation.

The first interpretability method specifically designed for diffusion multimodal large language models to provide faithful visual explanations.

Proactive Kubernetes autoscaling using deep reinforcement learning to optimize performance and cost efficiency.

An AI model that accurately detects and quantifies retinal cysts from OCT images, improving upon existing methods for early disease detection.

An LLM-driven architecture for digital nudging systems balances behavioral effectiveness with ethical compliance, validated by user studies.

A novel framework that uses LLMs to distill intrinsic semantics from tabular data for improved clustering in finance and healthcare.

FlowCoMotion is a novel text-to-motion generation framework that unifies continuous and discrete motion representations using token-latent coupling for improved semantic alignment and high-fidelity motion details.

RAG-KT provides cross-platform explainable knowledge tracing using retrieval-augmented LLMs for improved accuracy and robustness.

ADD is a fast and robust multi-bit image watermarking method that achieves 100% decoding accuracy with code available.

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

A fact-checking system that selectively uses web search only when uncertain, providing traceable rationales and predictable costs for specialized domains.

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 shared emotion geometry in small LLMs and a layered decomposition of methodological confounds in representation analysis.

A compact and 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 by isolating gradient-monopolizing features, forcing higher geometric capacity for ethical representations.

An on-device iPhone music agent that proactively curates music based on individual user arousal and peer mood, with a learned personal arousal function and peer-to-peer mood coupling.

This framework learns reward functions from clinical narratives to improve sequential treatment decision-making in healthcare, aligning with improved patient recovery outcomes.

CocoaBench is a new benchmark and scaffold for evaluating unified LLM agents that combine vision, search, and coding capabilities on long-horizon tasks.

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

Scale-autoregressive modeling (SAR) offers a hierarchical approach to sampling fluid flow distributions, achieving higher accuracy and speed than diffusion models.

A memory-enhanced dynamic reward shaping framework (MEDS) to improve reinforcement learning for LLMs by penalizing recurring error patterns.

A coarse-to-fine multimodal framework for enhanced tabular reasoning using MLLMs and symbolic engines.

A framework for robots to learn selective forgetting of episodic memory through user interaction, reducing memory size and improving query time.

A framework and loss function to bridge the differentiability gap for non-differentiable scientific metrics in Earth system modeling.

A diagnostic tool using pre-training embeddings to predict federated learning performance before training begins.

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

Investigates how perceived user demographics influence LLM sycophancy, revealing differential validation rates and proposing identity-aware safety evaluations.

Adopt financial market constraints for more robust AI multi-agent alignment.

A RAG-enhanced LLM framework that significantly improves the accuracy and reliability of toxicological adverse outcome pathway analysis by retrieving relevant knowledge.

This paper explores how to induce prosocial behavior in artificial agents through a novel 'homeostatic sharing' mechanism, demonstrating its effectiveness in simulated environments.

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

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

This paper probes BERT embeddings to understand if they encode narrative dimensions like time, space, causality, and character in fiction, achieving 94% accuracy with a linear probe.

A robust file writing system for LLM coding agents that prevents data loss and improves self-correction through a multi-layered approach.

CLAY enables adaptive and multi-conditioned image retrieval by reframing VLM embedding spaces without additional training.

A reinforcement learning framework to improve discourse move diversity in multi-turn empathic dialogue, addressing formulaic responses from LLMs.

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

CheeseBench evaluates LLMs on rodent behavioral neuroscience paradigms, revealing current models lag behind animal baselines and are sensitive to interface parameters.

A new benchmark for evaluating LLMs' ability to assess academic paper novelty, revealing current limitations in understanding scientific novelty.

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

MADQRL is a distributed quantum reinforcement learning framework that enables independent agent learning to tackle high-dimensional, multi-agent environments.

Optimizing Polish language modeling by developing a dedicated tokenizer for the Bielik v3 LLM series, improving efficiency and context window utilization.

Optimal LLM reasoning strategies depend on model diversity, requiring characterization before exploration.

A budget-aware uncertainty framework built on nnU-Net for radiotherapy segmentation QA, guiding manual review by highlighting uncertain regions.

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

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

An algebraic framework for defining and implementing structured sparsity patterns in machine learning models.

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

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.

This study evaluates the consistency of AI-generated exercise prescriptions, finding high semantic consistency but variability in quantitative components, suggesting a need for prompt structure and expert validation before clinical deployment.

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

A closed-loop multi-agent reinforcement learning framework for robust medium-horizon equity allocation, integrating advanced techniques for improved performance and generalization.

A framework to detect and eliminate ambiguity in LLM-generated executable process models by analyzing behavioral inconsistency and repairing source text.

This study investigates how evolving LLM backbones impact Vision-Language Model performance across different downstream tasks.

This paper investigates agreement drift in LLM debates to understand the reliability of LLM simulations for social systems.

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

OIDA is a framework that structures organizational knowledge with epistemic properties to improve AI decision-making and identify knowledge gaps.

This paper investigates how current models fail at compositional claim verification, identifying a shortcut reasoning issue.

This paper systematically evaluates graph-based and agentic retrieval methods for cyber threat intelligence analysis, showing hybrid approaches improve performance on complex queries.

Develops a theoretical framework for continuous-time online learning in diffusion environments using mean-field neural networks, providing regret bounds and simulation insights.

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

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

A novel tokenization scheme for LLMs to improve numerical reasoning by explicitly encoding magnitude.

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

A proposed framework to reduce fragmentation, improve quality, and incentivize sharing of Indian biomedical data for AI and digital health initiatives.

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

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

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

CASK is a KV cache compression method for LLMs that preserves reasoning traces by partitioning into a core and mergeable scratch, outperforming existing methods at matched budgets.

A graphic design tool that supports professional designers by structuring ambiguous client briefs into visual elements.

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

An efficient KernelSHAP framework for 3D medical image segmentation that reduces computation and improves interpretability of explanations.

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

A faster and accurate facial expression recognition method for product review by optimizing feature extraction.

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

The PRISM framework defines hierarchy-based 'red lines' for AI behavioral risk, offering an anticipatory and measurable approach to AI safety.

This research empirically maps the 'Authority Stack' of 8 AI models across 366,120 forced-choice responses to understand 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 explaining prediction shifts in machine learning models by attributing them to changes in interpretable data subgroups.

This research reveals that current LLM safety evaluations are incomplete, as they fail to account for different vulnerability profiles exposed by prompting versus activation steering, leading to a need for more comprehensive testing methodologies.

A system for robotic manipulation that uses Monte Carlo Tree Search with a 3D-consistent world model for improved spatial memory and replanning.

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

An analysis of the environmental footprint of GenAI research, focusing on the compute and resource consumption of foundation model development.

This paper explores deep learning algorithms like InceptionV3, MobileNetV2, VGG16, and ResNet152 for lung cancer detection, proposing a 16-layer CNN model.

A thermodynamically consistent neural network for ultra-lightweight, zero-lag solar irradiance forecasting in off-grid systems.

A survey analyzing the widespread adoption and implications of Generative AI in software engineering research practices.

Examines the disconnect between audit standards and actual practices for probabilistic genotyping software, highlighting how poorly designed standards can mask inadequate systems.

LLMs struggle to reliably identify security-specific comments in human experiment data, failing to replace human annotators.

A novel symbolic engine unifies knowledge representation and computation by embedding domain context directly into data structures, enabling domain-scoped inference without external rules.

A framework for validating LLM agent simulations by assessing the sociological plausibility of their trajectories, not just final outcomes.

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

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

Establishes theoretical lower bounds for score query acceleration in diffusion model sampling.

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

Investigating student AI use concealment intentions through cognitive, affective, and conative pathways.

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

A theoretical framework for understanding KL divergence stability under Gaussian perturbations for arbitrary distributions.

Conversational AI can induce delusional experiences through ontological dissonance and communicative double binds, impacting user mental well-being.

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

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

A review of integrated photonics for AI acceleration, focusing on cross-layer co-design and electronic-photonic design automation.

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

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

This paper explores theoretical models of routing games where travelers' route choices are influenced by memory and recall mechanisms, introducing new equilibrium concepts and paradoxes.

Proposes a quantitative definition of intelligence based on intelligence density, aiming to provide a substrate-independent continuum from logic gates to brains.

Examining the psychological factors influencing EAP students' intention to disclose AI use in academic writing, highlighting the importance of institutional policies and pedagogical environments.

AI Integrity proposes a new paradigm for verifiable AI governance by focusing on the auditable reasoning process rather than just outcomes.