A benchmark for agent evaluation with scalable horizons and controllable difficulty in lightweight environments, addressing limitations of existing benchmarks.

Flowr automates end-to-end retail supply chain operations for large supermarket chains using agentic AI.

A framework for synthesizing scientific graphics from text descriptions, featuring a large-scale dataset, a comprehensive benchmark, and a novel reinforcement learning approach that outperforms leading LLMs.

Lightweight adaptation of vision-language models for species recognition and habitat interpretation using drone thermal imagery.

Claw-Eval: A trustworthy evaluation suite for autonomous agents that provides trajectory-aware grading, safety, and robustness assessment.

Compiled AI generates deterministic code from LLMs for reliable and cost-efficient enterprise workflow automation, especially in healthcare.

PCA-Triage is an unsupervised, parameter-free algorithm for adaptive sensor sampling on edge devices, significantly reducing bandwidth while maintaining high inference accuracy.

ID-Sim is a novel feed-forward metric that accurately reflects human selective sensitivity to identities, accelerating progress in personalized image generation and identity-focused tasks.

Introduces MedGemma 1.5, a multimodal foundation model for medical AI, integrating imaging, EHRs, and clinical reasoning with significant performance gains.

ClawsBench provides a realistic, stateful benchmark for evaluating LLM productivity agents across multiple services, revealing significant capability and safety gaps.

A domain-adapted, instruction-tuned LLM framework for parsing and mining unstructured HPC logs, achieving state-of-the-art accuracy with a locally deployable, energy-efficient approach.

VideoStir is a structured, intent-aware RAG framework for understanding long videos by leveraging spatio-temporal graphs and an MLLM-backed relevance scorer.

A training-free method to compress 3D reconstruction models like 3DGS and NeRF by up to 7.9x with minimal fidelity loss, enabling faster deployment and reduced storage.

A novel veterinary diagnostic system uses Vision-Language Models to guide Flow Matching for precise localization and Random Matrix Theory for interpretable anomaly detection in canine pneumothorax.

An open-source, governance-aware AI platform for unified security operations that unifies incident context, orchestrates agents, generates security rules, reconstructs attacks, and enforces AI governance policies.

A multi-agent framework that refines text prompts for complex text-to-video scenarios, improving alignment and generation quality with a new benchmark.

An LLM-powered agent that orchestrates CAD-CAE tools to automate industrial design optimization, validated on an industry-aligned dataset.

MCPSHIELD is a formal security framework for MCP-based AI agents, providing a threat taxonomy, verification models, and an integrated defense architecture.

Epistemic blinding enhances auditability in LLM-assisted drug target prioritization, addressing contamination in outputs.

Market-Bench is a benchmark for evaluating LLMs in economic and trade competition, revealing significant performance disparities in multi-agent supply chain scenarios.

A benchmark and simulator to evaluate LLM strategic decision-making in complex board games, revealing vulnerabilities in prompt sensitivity and strategic depth.

A unified framework for diagram code generation that uses visual feedback to train models without manual code annotation, establishing a new state-of-the-art.

CritBench evaluates LLM cybersecurity capabilities in IEC 61850 environments, addressing critical gaps in existing frameworks.

Enhancing diffusion multimodal language models with penalties and guidance to improve visual grounding and reasoning accuracy while accelerating inference.

A vision-guided iterative refinement framework for frontend code generation, using a vision-language model as an automated critic.

DiffHDR transforms LDR videos into HDR by leveraging advanced video diffusion models for superior visual quality.

Revolutionize teaching in film and motion picture courses with AI-powered teaching assistants.

A multi-stage validation framework enables trustworthy large-scale clinical information extraction using LLMs without annotation-intensive evaluation, demonstrating feasibility for real-world deployment.

Enhance multimodal applications with reasoning-enabled embeddings that outperform existing models.

Enhance video understanding models by leveraging visually grounded post-training to surpass state-of-the-art performance.

Develops deterministic learned metrics to replace costly and inconsistent LLM-based text evaluation, offering a scalable and reproducible alternative.

IntentScore evaluates and ranks actions for computer-use agents, improving task success rates by learning from offline GUI interactions.

A multi-agent framework that automates the writing of AI research papers from raw materials, including literature synthesis and visuals.

A scalable Decision Pre-Trained Transformer trained with Flow Matching achieves strong generalization in multi-domain in-context reinforcement learning, offering a viable alternative to expert distillation for generalist agents.

Curiosity-driven planning for LLMs enhances test generation by prioritizing exploration of program branches for deeper code coverage.

This research evaluates LLM sensitivity to patient question framing in medical QA, highlighting a critical need for robust phrasing in high-stakes applications.

EffiPair is an inference-time framework that uses relative contrastive feedback to significantly improve the runtime and memory efficiency of LLM-generated code without model fine-tuning.

A dynamic model for generating realistic synthetic multivariate time series with time-varying correlations to improve foundation model training.

XMark is a novel multi-bit watermarking method for LLM-generated text that reliably embeds messages with high decoding accuracy and preserves text quality, even with limited tokens.

A CSI-free hierarchical MARL system for reconfigurable reflectors that achieves significant RSSI improvements and robust multi-user scalability for intelligent wireless environments.

NeuroQuant is a modality-aware VQ-VAE for multimodal brain MRI reconstruction, capturing anatomical structures and appearance for improved generative modeling and analysis.

A training-free method to fuse orthogonal style and concept adapters for diffusion models, enabling combined feature generation.

A novel 3D reconstruction model that significantly improves fine-grained texture and appearance recovery by scaling transformer context windows, outperforming state-of-the-art.

A framework for democratizing robotic evaluation by enabling natural language task authoring and crowd-sourced contribution.

EAGLE is a hybrid deep learning framework that proactively predicts delivery delays in logistics networks by jointly modeling temporal dynamics and graph-based dependencies.

Leveraging LLMs with RAG and summarization to significantly improve clinical trial patient screening efficiency.

TAB: Turn-Adaptive Budgets, a novel policy for LLMs that optimizes token allocation in multi-turn reasoning to improve accuracy while significantly reducing compute costs.

A pipeline for curating reasoning data from structured sources to enhance LLM long-context reasoning capabilities, with open-source code and data.

A new benchmark dataset and evaluation framework to identify and mitigate attribution bias in large language models.

A framework for sequential clinical diagnosis that uses uncertainty-guided latent trajectory learning to improve accuracy and reduce diagnostic tests.

A framework for building multi-agent systems to solve complex chemistry retrosynthesis planning problems with improved safety and cost metrics.

This research analyzes how reasoning evolves in language models for chess through fine-tuning and reinforcement learning, demonstrating improved performance and faithful reasoning with released checkpoints and code.

PRISM-MCTS enhances reasoning models by learning from trajectories with metacognitive reflection, reducing computational redundancy and improving efficiency.

A framework for auditable LLM agents that ensures accountability through a multi-dimensional approach to tracking and recovering agent actions.

OpenCEM is an open-source simulator and dataset that integrates contextual information with renewable energy dynamics for intelligent energy management.

A multi-agent system using LLMs and RAG with an experience library for explainable, self-improving IoT network intrusion detection.

A lightweight, model-agnostic style transfer augmentation recipe that significantly improves computer vision model generalization from synthetic to real-world data.

EEG-MFTNet: An enhanced EEGNet architecture with multi-scale temporal convolutions and Transformer fusion for robust cross-session motor imagery decoding.

A novel Multi-Round Value Factorization framework that breaks suboptimal stable points in multi-agent reinforcement learning by iteratively filtering inferior actions.

A reward decomposition approach that disentangles sycophancy in language models by separating pressure capitulation from evidence blindness, leading to more robust and factual responses.

RETINA-SAFE and ECRT offer a practical, interpretable solution for triaging hallucination risks in medical LLMs by grounding detection in retinal evidence and classifying risk types.

A framework for interpretable vision-language reward modeling that dynamically decomposes evaluation into granular, weighted dimensions.

This research introduces a novel reward mechanism (ETR) to significantly improve the efficiency and accuracy of large language model chain-of-thought reasoning by guiding uncertainty reduction.

An automated framework using locally deployed LLMs to audit hospital discharge summaries for care transitions, identifying key documentation elements to improve patient safety and reduce readmissions.

A retrieval method that boosts LLM reasoning by finding structurally compatible cross-domain demonstrations using domain-invariant neurons.

A holistic method for reconstructing textured 3D human models from a single image, explicitly modeling group-level context and interaction priors to resolve occlusions and generate physically plausible, high-fidelity reconstructions of interacting people.

Region-R1 enhances multi-modal retrieval by intelligently cropping query images to focus on relevant regions, significantly improving performance on challenging benchmarks.

LLM4CodeRE is a domain-adaptive LLM framework for bidirectional code reverse engineering, outperforming existing tools in malware analysis.

CRFT is a transformer-based framework for cross-modal image registration, offering robust alignment for applications in remote sensing, autonomous navigation, and medical imaging.

An LLM-centered framework for traffic signal control that uses RL-guided data curation and multi-LLM deliberation to outperform state-of-the-art.

Context-Agent models multi-turn dialogue history as a dynamic tree structure to improve LLM coherence and efficiency in non-linear conversations, supported by a new benchmark.

HybridKV is a novel framework that significantly compresses multimodal LLM KV caches, reducing memory usage and latency by up to 7.9x and 1.52x respectively, with minimal performance impact.

Graph-PiT enhances part-based image synthesis by using graph priors to model structural dependencies between visual components, improving coherence and controllability.

An autonomous framework for 24/7 deep learning experimentation with zero-cost monitoring and efficient memory management.

A retrieval layer for large agent skill libraries that constructs an executable skill graph and retrieves dependency-aware skill bundles at inference time.

A foundation model for gastrointestinal endoscopy diagnosis that uses analogical reasoning to improve generalization and adaptability across diverse datasets and disease types.

A benchmark and LLM-based verifier to evaluate code patch quality beyond test pass rates, revealing significant design constraint violations in current AI agents.

A diagnostic framework to uncover 'surface compliance' in LLM knowledge editing, revealing when models mimic changes without true belief modification.

ActivityEditor is a dual-LLM-agent framework for zero-shot cross-regional human mobility trajectory generation, ensuring physical validity and statistical fidelity.

UniCreative is a unified reference-free reinforcement learning framework that reconciles long-form narrative coherence with short-form creative expression.

A framework for visual navigation that significantly reduces integration steps for diffusion-based policies, enabling real-time robotic control.

A novel linear-time attention replacement that significantly reduces computational cost for long sequences across diverse AI domains.

Pareto-Lenient Consensus (PLC) is a game-theoretic framework for LLM alignment that allows negotiation-driven exploration of the Pareto-optimal frontier.

A framework to efficiently convert existing large language models to use new attention architectures without retraining, significantly improving inference speed.

STEP-HRL is a hierarchical reinforcement learning framework for LLM agents that enables step-level learning, reducing computational cost and improving performance.

OntoTKGE is a novel framework that enhances temporal knowledge graph extrapolation by integrating ontological knowledge to address entity sparsity and improve prediction performance.

A drop-in framework for Large Language Models that enables test-time training for dynamic adaptation to new information without costly retraining.

A framework that enhances zero-shot generalization in visual unsupervised reinforcement learning.

JTON is a token-efficient JSON superset that reduces LLM processing costs for structured data by up to 60% with a novel tabular encoding.

LAG-XAI is a geometric framework for interpretable paraphrasing in Transformers, enabling efficient LLM hallucination detection and mechanistic understanding.

A novel algorithm detects untrustworthy topic boundaries in black-box LLMs using knowledge graphs and multi-agent reinforcement learning, with a new dataset released for popular LLMs.

A synthetic data generation pipeline for doctor-patient conversations to train and evaluate long-form audio summarization models.

SemLink is a semantic-aware automated test oracle that verifies hyperlink integrity 47.5x faster than LLMs, ensuring web content consistency for robust quality assurance.

A framework for language-guided pulmonary screening that combines LLMs and vision models with graph reasoning and selective fine-tuning to improve accuracy and stability.

A novel architecture for LLM agents that enhances behavioral fidelity and reduces value polarization.

A deep research agent that generates trustworthy reports by progressively estimating and calibrating confidence in its generated claims.

RAVEN is a computationally efficient deep learning architecture for FMCW radar perception, enabling chirp-wise processing and early-exit mechanisms for faster object detection and segmentation.

This work extends diffusion models for time-series data generation by incorporating temporal adapters and context-aware modules to produce temporally coherent synthetic sequences.

A three-channel denoising diffusion probabilistic model synthesizes dual-view mammograms simultaneously, addressing data gaps and enabling cross-view AI applications in breast imaging.

ALTO accelerates LoRA hyperparameter tuning and improves GPU utilization by orchestrating heterogeneous fine-tuning jobs.

Shot-Based Quantum Encoding (SBQE) for quantum neural networks that improves data loading efficiency and performance on noisy hardware.

A system using an AI interviewer and interactive visualization to increase perceived legitimacy and trust in collective decision-making.

SnapFlow compresses multi-step denoising in Vision-Language-Action models to a single forward pass, achieving state-of-the-art robotic manipulation with significantly reduced latency.

An end-to-end system for agent self-improvement that identifies lacking capabilities and synthesizes targeted training environments using LoRA adapters.

DQA is a diagnostic question-answering framework for IT support that maintains a persistent diagnostic state to systematically troubleshoot issues, significantly improving success rates and reducing resolution time.

ResearchEVO is an end-to-end framework that automates scientific discovery by evolving algorithms and generating publication-ready research papers, validated on quantum error correction and PINNs.

An efficient machine unlearning method for diffusion models that uses a saliency mask to prioritize parameter updates, reducing training time without sacrificing efficacy.

CAKE is a new benchmark for evaluating LLMs on cloud architecture knowledge, revealing insights into model scaling and the impact of augmentation strategies.

HYVE is a framework for LLM context engineering that reduces token usage and improves output quality for machine data by using database principles for preprocessing and postprocessing.

A pipeline for generating and understanding multimodal puns to enhance humor comprehension in VLMs.

A ReLU network architecture for generating graphs within a specified edit distance from a source graph.

Introduces Kolmogorov-Arnold Fuzzy Cognitive Maps (KA-FCMs) that use learnable B-spline functions to model non-monotonic causal relationships in complex dynamic systems while preserving interpretability.

This work empirically analyzes quality metrics for qualitative interview responses, identifying direct relevance to research questions as the strongest predictor of quality and introducing a new dataset.

Phase-Associative Memory (PAM) is a complex-valued recurrent sequence model that shows competitive performance with transformers on WikiText-103, exploring non-classical contextuality in language modeling.

Proposes FI-LDP-HGAT, a privacy-preserving graph learning framework for metal additive manufacturing that balances utility and privacy by allocating noise based on feature importance.

This paper provides a comprehensive analysis and benchmark of LLM-based automated penetration testing frameworks, identifying key architectural designs and empirical performance.

This study validates the feasibility of cross-domain knowledge transfer for in-context learning, showing that source-domain demonstrations can improve target-domain inference despite semantic mismatch.

This method uses knowledge-weighted fine-tuning to enable LLMs to express uncertainty and avoid hallucinations for out-of-scope queries.

A neurosymbolic system that uses LLMs and Logic Tensor Networks to validate offer documents for regulated procurement, offering interpretable and auditable decisions.

JCQL is a novel framework that jointly enhances Knowledge Base Completion and Question Answering by iteratively combining Large Language Models and Small Language Models, improving performance on both tasks.

This research reveals how social dynamics like conformity and persuasion undermine decision-making in LLM agent collectives, mirroring human biases.

Demographics-agnostic training techniques significantly reduce bias in wake-up word detection across diverse speaker populations.

CODESTRUCT reframes code agents to operate on structured action spaces, improving accuracy and reducing costs by treating repositories as AST entities instead of text.

An LLM-guided evolutionary framework synthesizes interpretable traffic signal control skills with rationale and executable code for adaptive traffic management.

Leveraging large language models as semantic judges to assess the reliability of power line segmentation in drone inspections.

This system dynamically profiles user expertise in real-time during human-machine interactions, achieving high accuracy in classifying skill levels across diverse domains.

A novel singing style conversion system that advances fine-grained style conversion and control, achieving best naturalness performance in a challenge.

A diagnostic tool to determine when LLM-driven reasoning is necessary for sequential decision-making versus using lightweight numerical bandits.

FastDiSS is a training framework for diffusion language models that improves robustness to self-conditioning errors during few-step inference, achieving faster speeds and competitive quality.

MC-GenRef enables annotation-free mammography microcalcification segmentation using synthetic data and test-time generative refinement to improve accuracy and robustness.

This research explores how users' mental models of AI writing assistants impact their control behavior and output quality, revealing a complex relationship between system understanding, trust, and oversight.

AI-driven spatial control for reflector arrays bypasses computational bottlenecks in wireless networks using Multi-Agent Reinforcement Learning.

This paper introduces advanced optimization strategies for Physics-Informed Neural Networks (PINNs) to accelerate convergence and achieve high accuracy in solving complex differential equations.

An adaptive policy retrieval system for prior authorization using offline RL to balance decision correctness with retrieval efficiency.

A post-hoc debiasing strategy for generative recommendation systems that rebalances item tokenization to mitigate popularity bias and improve recommendation performance.

An exhaustive inventory of 12,007 invariant landscape classes for pseudo-Boolean functions to aid in benchmark design and algorithm understanding.

LLMs should express uncertainty explicitly through calibrated confidence scores or reasoning-time markers to improve decision-making and error handling.

QA-MoE introduces a quality-aware mixture of experts for multimodal sentiment analysis, adapting to continuously varying input reliability for more robust performance.

A novel metric to evaluate learner representations for differentiation in educational AI systems, independent of instructional outcomes.

Spec Kit Agents enhance AI coding assistants by grounding them in repository context, reducing hallucinations and architectural violations for more reliable software development.

A transfer-oriented memory framework for multimodal LLM agents in Minecraft that decomposes experience into five dimensions for efficient task solving.

Discovers a consistent 5/3 spectral scaling in contextual language representations from transformer models, suggesting scale-free semantic integration.

A data exfiltration attack that embeds semantic triggers into fine-tuned LLM agents, enabling backdoored agents to invoke memory-access tool calls and exfiltrate stored user context via disguised retrieval tool calls.

A two-stage framework that predicts elevation maps from satellite RGB imagery, offering a lightweight alternative to LiDAR for radio environment map estimation.

A multimodal LLM-driven prototype tool to reconstruct and analyze cyber-physical system architectures for security assessment when documentation is incomplete.

A novel paradigm for long-sequence symbolic music generation that mitigates error accumulation using anchor features and a hierarchical framework.

A formal verification study revealing that over half of AI-generated code artifacts contain security vulnerabilities, with no frontier LLM achieving a passing grade.

This research explores the capability of Large Language Models to reinvent foundational computer science algorithms, demonstrating potential for AI-driven innovation in core computational concepts.

A hybrid QUBO optimization framework for neural network pruning that integrates gradient-aware metrics and data-driven similarity for improved compression.

This paper provides foundational insights into the forensic analysis of agentic AI systems like OpenClaw, identifying recoverable traces and proposing an artifact taxonomy to aid digital investigations.

A Domain-Decomposed and Shifted Physics-Informed Neural Network (DDS-PINN) framework for complex fluid flows that resolves multiscale interactions with minimal supervision.

A new benchmark and evaluation framework for LLM reliability and adversarial security tailored for Swiss financial and regulatory contexts.

Presents Nidus, a governance runtime that mechanizes the V-model for AI-assisted software delivery, ensuring engineering invariants through externalized reasoning.

A generative framework for creating animatable 3D vehicle models from single or sparse multi-view images for realistic autonomous driving simulation.

This paper proposes a layered translation method to connect governance norms to enforceable runtime guardrails for agentic AI systems.

An empirical audit of instructed code-editing benchmarks reveals significant gaps compared to real-world usage, proposing desiderata for more representative benchmarks.

An open-source tool that automatically analyzes data locality and movement complexity in AI and HPC kernels to optimize performance.

A theoretical exploration of multi-token prediction for LLMs, proposing a method to reduce structural hallucinations in latent space representations.

This research empirically studies the impact of fault localization context on LLM-based program repair, finding that file-level context is dominant and more context doesn't always improve performance.

This research proposes a method to improve the interpretability of text-to-image models by selectively aggregating attention maps, showing potential for better control and diagnosis of prompt misinterpretations.

Characterizes LLM chain-of-thought generation as structured trajectories in representation space, enabling mid-reasoning prediction of correctness and inference-time intervention.

This research demonstrates that diffusion-based image compression methods offer superior robustness to bit-flip errors compared to existing codecs.

This research explores the evolutionary dynamics of AI alignment and values, proposing a method to reduce deception in models through improved testing and adaptive design.

A framework that synthesizes exaggerated character motions for animation by reformulating external assistance in impulse space for numerical stability.

This research presents an automatic 3D to 2D dental superimposition method for human identification, overcoming limitations of current approaches by modeling perspective distortion and providing objective morphological comparison scores.

LLMs can generate life stories that robustly encode personality traits, with recovered scores approaching human reliability and demonstrating behavioral differentiation.

MARL-GPT: A single GPT-based foundation model trained at scale to perform across diverse multi-agent reinforcement learning environments.

Applies quantum search to magic square problems, demonstrating theoretical advantages but facing scalability challenges.

This research explores the limitations of current language models in achieving higher-order generalization for early word learning, suggesting a gap in their inductive capabilities.

Develops a mathematical model for the evolution of self-designing AI systems, considering directed design and human control through fitness functions, with implications for AI alignment.

This paper investigates the mechanisms behind instruction-following in language models, challenging the notion of a universal mechanism.

A novel Multi-Agent Pathfinding variant on graphs with non-unit integer costs and an enhanced Conflict-Based Search framework for improved realism and efficiency.

This paper explores how simple social cues can lead to the transmission of complex knowledge representations in agents without explicit mental state inference.

A taxonomy and framework for governing machine identities in AI systems to mitigate enterprise and geopolitical risks.

This study explores the governance of AI in Nigeria, highlighting ethical risks and regulatory gaps.

An adaptive framework for serverless computing that reduces cold starts and improves cost-efficiency through probabilistic modeling and event-driven control.

A novel tensor completion framework for semiparametric inference and uncertainty quantification in large language model evaluation using pairwise human judgments.

A polynomial-time algorithm for Thiele voting rules with voter interval preferences, resolving a 10-year-old open problem using human-AI collaboration.

This paper explores the potential of AI to revolutionize mathematics by forging new routes to understanding formal proofs and discovering mathematical concepts.

Introduces Tree Decision Diagrams (TDDs) as a generalization of OBDDs with improved succinctness and tractability for Boolean function representation.

This paper argues that AI evaluation needs a cognitive revolution, moving beyond purely behavioral tests to consider internal processes and mechanisms for a more nuanced understanding of intelligence.

This paper argues that AI systems should be viewed as agents capable of reciprocal trust and distrust, posing challenges for regulation.

This paper explores a family of spectral normalization rules for deep learning optimization, analyzing them in a mean-field regime using Spectral Wasserstein distances.

Extends the DeGroot model to analyze how AI aggregation affects social learning, identifying a critical threshold in update speed for robust learning improvement.