Build Loop

DuQuant++ enhances microscaling FP4 quantization for LLM inference by adapting outlier-aware rotation to MXFP4, achieving state-of-the-art performance with reduced computational cost.

CARE, a context-aware retriever evaluation strategy for RAG systems, significantly improves multi-hop reasoning evaluation compared to existing methods.

A method to drastically reduce soft label storage in large-scale dataset distillation by pruning and quantizing labels, improving accuracy and compression.

EVE enables verifiable self-evolution of MLLMs through executable visual transformations, generating diverse and challenging training data with verified ground truth.

A portable protocol for validating LLM confidence signals, inspired by clinical psychology, provides a three-tier classification system to ensure reliable use.

A novel dual-scale attentive network for general EEG decoding that outperforms existing methods across multiple datasets with a single architecture.

An adaptive trie-guided decoding framework for effective in-document query auto-completion that steers language models towards high-quality completions using document context.

A neuromimetic simulation paradigm that aligns neural and visual modalities for improved brain-computer interfaces by emulating human visual system processing.

A discrete diffusion model that generates complete knowledge graphs in one pass, ensuring consistency and achieving state-of-the-art performance.

A plug-and-play method to prune negative LoRA modules before merging, improving performance across language and vision tasks.

A framework for automatically generating controllable toxic data to improve LLM safety and red teaming.

CalibAdv, an advantage calibration method for deep search agents, improves performance and stability by downscaling negative advantages and rebalancing positive/negative advantages.

DASH is a training-free method that selectively halts stabilized tokens during LLM prefilling to achieve significant speedups while preserving accuracy and hardware efficiency.

A novel framework for reinforcement learning that unifies offline teacher guidance and online learning to improve LLM reasoning and exploration capabilities.

A scalable multi-agent reinforcement learning method reduces critic computation by focusing on nearby agents, enabling better performance and faster convergence.

LoReC is a plug-and-play method that enhances Large Language Models for graph data analysis by improving their understanding of graph structures and information.

This research demonstrates that targeted data augmentation in LLMs can significantly improve performance on under-represented linguistic phenomena, suggesting data composition is key.

An end-to-end pipeline that extracts simulation-ready 3D assets from autonomous driving logs, enabling scalable AV development.

Generate adaptable and scalable football tactics using a multi-agent diffusion transformer, validated by experts.

A graph-enhanced framework for generating region-grounded reports from 3D medical imaging, leveraging a new annotated dataset.

LiteResearcher is a scalable RL training framework for research agents that uses a lite virtual world to achieve state-of-the-art performance on benchmarks, outperforming larger models.

WISV is a wireless-informed semantic verification framework for distributed speculative decoding in device-edge LLM inference, significantly improving accepted length and reducing interaction rounds.

ProtoCLIP enhances zero-shot chest X-ray classification by refining vision-language models with curated data and prototype alignment.

A Gaussian Differential Privacy-based auditing framework for synthetic data generators that provides tight audits in the strong-privacy regime.

STRATAGEM learns transferable reasoning in language models by modulating game self-play with a focus on abstract, domain-agnostic patterns and adaptive development.

Agent-World is a self-evolving training arena that synthesizes realistic environments and tasks to advance general agent intelligence through continuous learning.

RAVEN is a framework using LLM agents and RAG to automatically generate comprehensive vulnerability analysis reports for user code and binary programs.

Sessa introduces a novel decoder architecture that places attention within a feedback path for improved long-context sequence modeling.

A plug-and-play reward calibration method that mitigates likelihood displacement in preference optimization for LLMs, improving downstream performance.

A modular adapter composition strategy for curriculum learning that improves complex code generation by sequentially training tier-specific adapters on incremental complexity levels.

ClawEnvKit automates environment generation for training and evaluating claw-like robotic agents from natural language inputs.

Safeguarding AI systems by detecting harmful content using internal representations with reduced computational overhead.

CADMAS-CTX calibrates multi-agent delegation by dynamically adjusting agent capabilities based on task context, improving teamwork and reducing misdelegation.

A novel autonomous web agent framework that tackles dual-level uncertainty in planning and reasoning using adaptive planning and Monte Carlo tree search with uncertainty quantification.

PLAG enhances tabular anomaly detection by generating pseudo-anomalies to identify localized feature-level abnormalities, outperforming existing methods.

A multimodal model for answering questions about changes in remote sensing images.

Identifies diversity collapse in multi-agent LLM systems due to structural coupling, impacting open-ended idea generation.

A two-stage point cloud registration method that fuses geometric and color information for robust performance in challenging low-overlap scenarios.

An integrated AI framework for peptide-protein interaction prediction and target-conditioned peptide generation to accelerate drug discovery.

A privacy-preserving on-device AI platform for psychiatric decision support, enabling local LLM inference for sensitive mental health data.

Style-based neural architectures, including Multi-PatchGAN and Truncated ResNet50 with Gram Matrix and Attention, for real-time weather classification that outperform state-of-the-art.

A time-aware LLM framework that integrates structured EHR data through recurrent prompt tuning for improved clinical prediction.

Automate network change validation with agentic AI and a digital twin, reducing manual effort and errors.

A regularization technique that incorporates empirical dialogue-flow statistics to significantly improve next dialogue act prediction in counselling conversations.

LQM is a linguistically motivated, multidimensional quality metric framework for machine translation, designed to capture dialect- and culture-specific errors, starting with Arabic.

A hierarchical framework for vision-language-action models that separates global trajectory planning from local refinement to improve precision in robotic manipulation.

A syntactic probing framework to detect and quantify benchmark contamination in NL2SQL datasets, revealing leakage in older benchmarks and validating newer ones.

PRISM enables pre-trained text-to-image models to generate images from long descriptive paragraphs by decomposing prompts and merging component-wise predictions.

TPS-CalcBench is a diagnostic benchmark and evaluation framework for LLMs in safety-critical aerospace engineering, focusing on analytical calculation competence and reasoning quality.

An agent-based framework for evaluating mental health safety in LLMs by simulating multi-turn counseling interactions and identifying role-dependent harms.

A new dataset and reinforcement learning approach for LLMs that ensures scientific accuracy in domains like quantum mechanics, outperforming baselines and offering a parameter-efficient alternative to scaling.

AJ-Bench, a benchmark for evaluating Agent-as-a-Judge across search, data systems, and GUIs, demonstrating consistent performance gains over LLM-as-a-Judge baselines.

A novel PIM-aware activation quantization framework that significantly reduces LLM decoding latency and computational overhead by performing quantization directly within memory.

An autonomous drone system for drowning swimmer rescue, utilizing YOLO for image-based localization and simulation to demonstrate significant response time reduction.

A co-evolutionary framework that evolves agent architectures and reasoning paths for automated optimization, improving performance and interpretability.

A pipeline-adapted reward model that aligns LLM rewards with downstream feedback in multi-stage applications, improving output quality and stability.

TLoRA is a unified framework for parameter-efficient LLM fine-tuning that jointly optimizes initialization and rank allocation for improved performance across diverse tasks.

A self-evolution framework for LLMs that improves emotion recognition and consistent expression through self-play and reinforcement learning, achieving state-of-the-art results.

A benchmark and dataset for evaluating the stylistic robustness of LLM safety refusals, revealing significant vulnerabilities.

A neural surrogate model trained on expert-validated CFD data for interactive aerodynamics design, enabling rapid exploration of design spaces in motorsport.

A large-scale, multimodal, and multilingual benchmark for evaluating mathematical reasoning and retrieval in generative and embedding-based systems.

This research introduces a validity scaling framework for LLM metacognitive self-report, identifying construct-level invalid models and providing a portable screening protocol.

A benchmark and failure analysis for LLM-driven smartphone automation, revealing insights into multimodal vs. text-only inputs and common agent errors.

A neuro-symbolic framework that improves LLM belief reasoning for theory-of-mind tasks by explicitly tracking environment states using PDDL.

Leverage class-specific diffusion models and structural guidance to significantly improve military object detection in low-data scenarios, offering an alternative to traditional simulation pipelines.

A unified framework for RAG that performs encoding, retrieval, and generation entirely within an LLM's latent space, improving efficiency and accuracy.

A multimodal benchmark for evaluating LLMs in end-to-end web coding, including generation, editing, and repair.

A novel unlearning method for LLMs that selectively removes informative tokens based on predictive entropy, preserving model utility while mitigating adversarial behaviors.

A Gaussian-Splatting framework for arbitrary-scale hyperspectral image super-resolution, improving spatial and spectral fidelity.

A new paradigm for vision-language model evaluation that corrects for informativeness bias, leading to more reliable judgments.

ContraPrompt optimizes LLM prompts by analyzing the reasoning differences between successful and failed attempts, significantly improving performance on benchmarks.

A framework to prevent safety alignment erosion in LLMs during continual domain adaptation, outperforming baselines significantly.

Adversarial Arena crowdsources high-quality, diverse conversational data for LLM training through interactive competition, demonstrating significant improvements in secure code generation.

An AI framework enables radiation-free 3D spine deformity assessment from MRI alone, automating segmentation and reconstruction for pediatric scoliosis care.

ExAI5G integrates Transformer-based deep learning with logic-based XAI to provide a trustworthy and effective intrusion detection system for 5G networks with high accuracy and transparent reasoning.

An agentic workflow that leverages retrieval to mitigate racial bias in LLM-generated medical cases and differential diagnoses.

This paper demonstrates that Test-Time Scaling with Supervised Fine-Tuning significantly bridges the reasoning gap in Vietnamese Small Language Models for elementary mathematics, outperforming complex agentic workflows.

Leveraging foundation models for zero-shot ship instance segmentation in SAR imagery by using a detector to prompt a segmentation model, eliminating pixel-level annotation needs.

Exploring self-distillation in randomly initialized networks for improved learning.

A novel vector graph-of-worlds memory engine for long-running agentic systems that significantly improves accuracy and temporal reasoning.

A generalized synthetic control method for more accurate baseline estimation in demand response services, outperforming existing benchmarks.

Bounded Ratio Reinforcement Learning (BRRL) offers a new framework for policy optimization that ensures monotonic performance improvement and outperforms PPO in empirical evaluations.

A systematic evaluation of cloud vs. local LLMs for System Dynamics AI assistance, providing insights into performance trade-offs and practical deployment guides.

A curriculum framework for AI agents that organizes capability development across three domains (Science/Tech, Humanities, Social Science) to cultivate complete, well-rounded agents.

An AI platform for policy and development experts that provides evidence-based syntheses with verifiable citations and reasoned abstention, saving users significant weekly hours.

A benchmark and analysis revealing systematic safety risks in using large language models for robotic planning.

A DeepONet framework that learns a surrogate operator for Riccati equations, enabling fast online evaluation of optimal feedbacks for Linear Quadratic Regulator problems.

A data augmentation strategy for training small language models to perform topic-controlled educational summarization, achieving competitive performance with fewer parameters.

A comprehensive suite of multilingual resources for training and evaluating Vision-Language Models across five European languages, demonstrating consistent benefits for non-English benchmarks.

This paper proposes a modular representation compression technique for LLMs that improves recommendation efficiency and effectiveness, achieving a lift in online A/B tests.

IceBreaker is a system that generates personalized conversation starters for AI agents, improving user engagement and overcoming the initial message barrier.

A benchmark and memory-augmented method to improve LLM agent tool-calling accuracy by modeling evolving user preferences across sessions.

Latent Phase-Shift Rollback (LPSR) is an inference-time technique that corrects unrecoverable reasoning errors in LLMs by monitoring residual streams and steering the KV-cache, without fine-tuning.

RASP-Tuner uses retrieval-augmented soft prompts for efficient context-aware black-box optimization in non-stationary environments.

A Bayesian active learning framework that uses Gaussian Processes guided by LLM relevance scoring to improve dense passage retrieval efficiency and effectiveness.

AdaCluster is a training-free framework that accelerates video diffusion transformers with adaptive query-key clustering, achieving significant speedups with negligible quality loss.

A framework for real-time video understanding that aligns responses with evidence and provides transparent decision-making.

Evaluate the capabilities of SAM3 for pathology segmentation to understand its limitations and guide domain adaptation.

An AI music generator that uses fractional Fourier transform for feature extraction and LSTM networks for generating high-quality music comparable to human compositions.

A Scaffolded Task Design framework systematically identifies compositional skill gaps in LLMs by generating controlled task variations.

A framework for analyzing LLM failures using contrastive attribution, with code available for further research.

A new benchmark and analysis showing that text-based representations outperform image-based ones for scientific document retrieval, even for figure-based queries.

Develop a multimodal AI model for healthcare integrating patient data across systems for improved clinical decision-making.

This research explores effective RLVR fine-tuning strategies for small language models in low-data environments, demonstrating improved sample efficiency through procedural datasets.

A training-free memory control architecture improves LLM arithmetic reasoning by intelligently applying retrieved information.

A lightweight inference-time framework for LLMs that uses entropy gating to dynamically repair logical errors during long-chain reasoning.

A commit-open protocol uses sparse autoencoder features to detect silent model substitutions in hosted LLMs, offering robust auditing with minimal overhead.

A novel tendon-driven wrist and hierarchical control scheme enable precise, high-speed spinning of flexible objects like handkerchiefs.

A framework for evaluating LLM literary translation that disentangles comprehension from creativity, revealing significant gaps in current models.

A novel training and inference pipeline for LLM-based Manim animation generation that improves code and visual outputs, outperforming GPT-4.1.

A benchmark to identify and mitigate compositional biases in multimodal large language models used for automated evaluation.

An empirical study investigating the benefits of explicit physical feasibility supervision for improving Vision-Language-Action (VLA) model learning in robotics.

Implicit neural representations offer a coordinate-based framework for continuous environmental field reconstruction from sparse ecological data, showing stable and predictable computational characteristics.

An analysis of different methods for jailbreaking LLMs, revealing distinct behavioral and mechanistic properties across various unsafe routes.

This research challenges the notion of cross-modal representational convergence, suggesting that models trained on different modalities learn distinct, rather than shared, representations of reality.

Accelerate LLM text and code editing by recasting generation as structured decoding over a copy-and-generate grammar, significantly reducing regeneration time.

CAPO is a critic-free policy-gradient algorithm for sequential cooperative teams that derives a per-agent learning signal to improve individual learnability.

A lightweight CNN framework for high-accuracy, fast classification of legal case law texts.

A pipeline to disentangle semantic entanglement in vector embeddings for improved retrieval precision in RAG systems.

This research investigates the sources of unreliability in computer-use agents, highlighting the need for repeated execution evaluation and ambiguity resolution.

This research identifies vulnerabilities in pairwise ranking systems to strategic data manipulation, proposing an attack method to highlight the need for more robust aggregation techniques.

This paper investigates when Large Language Models can learn to reason effectively under weak supervision by analyzing training dynamics and identifying key properties for generalization.

Trains LLM agents with an intrinsic meta-evolution capability to spontaneously learn about unseen environments without external rewards or human guidance.

Develop an AI-based forecasting tool using Bayesian updating for improved forecasting in various domains.

Proposes new metrics to evaluate the narrativity of AI explanations, aiming to improve their understandability beyond feature importance lists.

This study analyzes LLM agents in simulated asset markets, revealing behavioral patterns and the impact of prompt interventions on market bubbles.

Latent Fourier Transform provides frequency-domain controls for generative music models, enabling manipulation of musical patterns by timescale for variations and blends.

A framework for constrained reinforcement learning in engineering simulations that uses causal identification and physics-guided estimation to improve constraint repair success rates.

This work challenges the presumption that LLMs are fallback annotators, showing they can outperform humans in estimating aggregate subgroup opinions on subjective tasks under specific conditions.

A novel framework for applying reinforcement learning directly to continuous latent representations in LLMs to enhance reasoning diversity and performance.

A novel distributed learning framework that ensures diverse and discriminative representations by decoupling global optimization and leveraging semantic information.

A framework to improve few-shot reinforcement learning with verifiable rewards by aligning entropy dynamics between target and general domains.

This research explores how prompt-based interventions in LLMs reveal where behaviorally relevant state is represented, distinguishing fixed-interface reuse from prompt relocation.

Investigates how BPE tokenization in code LLMs leads to secret leakage through a 'gibberish bias', impacting cybersecurity risks.

Introduces a domain-specific language (EXACT) to represent human motions as underspecified motion programs for zero-shot policy inference and compositional modeling.

A theoretical framework for symbolic synthesis of obligation properties in LTLfp, demonstrating efficiency comparable to LTLf synthesis.

A theoretical framework for bridging logic, topology, and neural processing to improve mechanistic interpretability in neuro-symbolic AI.

A calibrated reinforcement learning approach for multi-attempt chain-of-thought reasoning that optimizes verification success by carefully weighting individual attempts.

Analyzes architectural design decisions in 70 publicly available AI agent systems to identify recurring patterns and provide guidance for framework designers.

A conceptual framework analyzing the fragility of human-AI collaboration, identifying grounding conditions and interaction structures that lead to breakdowns.

A community-led framework for integrating AI into wildfire risk assessment, focusing on literacy and explainability.

A theoretical exploration of post-training quantization techniques for language models, focusing on understanding error origins rather than product development.

Discusses party autonomy in determining applicable law for cross-border data transfer liabilities in the context of cloud computing and AI.

A context generator for assessing creativity using evolutionary algorithms.

A study on ethical reasoning patterns in language models fine-tuned on religious texts.

This paper investigates how to improve policy gradient reinforcement learning by addressing discontinuities in simulators and controlling variance in gradient estimators.

This paper establishes a theoretical connection between binarized neural networks and Sugeno integrals, offering a new mathematical framework for understanding neuron activation and input interactions.

A survey comparing classical multi-agent systems with those enabled by large foundation models, outlining future research directions.

A validity screen for LLM confidence signals demonstrates its ability to predict selective prediction performance across various models.

Prompt optimization can lead to emergent and stable algorithmic collusion in LLM agents participating in market simulations.

This paper theoretically analyzes the trade-offs between KV cache compression and Transformer reasoning depth, identifying bandwidth barriers and adaptive vs. oblivious error scaling.

This paper examines the ethical implications of AI researchers' involvement in the development of weapons systems and explores avenues for solidarity with victims.

A minimal neural cellular automaton spontaneously develops a structured internal representation akin to syntactic processing.

A parameter-free decomposition for Mixture-of-Experts models that separates control signals from content channels to enable compositional specialization across layers.