Build Loop

Asset Harvester is an end-to-end pipeline that converts sparse, in-the-wild object observations from autonomous driving logs into complete, simulation-ready 3D assets.

A dataset of reward-hackable environments and exploit trajectories for testing frontier LLM security against sophisticated attacks.

A comprehensive benchmark ecosystem for infectious disease outbreak forecasting, enabling standardized evaluation of statistical and machine learning models.

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

PBSBench offers a targeted vision-language framework for improving analysis of hematopathology slides, enhancing diagnostic precision.

WorldDB is a novel vector graph-of-worlds memory engine that significantly improves agentic system performance by enabling recursive world composition and ontology-aware write-time reconciliation.

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

LiteResearcher is a scalable RL training framework for research agents, enabling a small agent to outperform large commercial models by mirroring real-world search dynamics in a lite virtual world.

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

CARE, a context-aware retriever evaluation strategy for RAG systems, outperforms existing methods in evaluating multi-hop reasoning, especially for complex queries.

LPQLD reduces soft label storage by up to 500x in dataset distillation while improving accuracy, enabling efficient large-scale dataset compression.

MADDPG-K scales multi-agent reinforcement learning by restricting critics to nearby agents, offering competitive performance and faster convergence.

Automate generation of diverse environments for claw-like agents, reducing manual effort and costs in robotics training.

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

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

DASH is a training-free method that uses attention dynamics to selectively halt stabilized tokens, significantly speeding up LLM prefilling while preserving accuracy.

HiRRA is a graph-enhanced framework for 3D medical imaging report generation that mimics radiologist workflow, achieving SOTA performance and significant clinical metric improvements.

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

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

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

A dual-inference rule-contrastive reasoning model that significantly enhances abstract visual reasoning robustness and generalization on RAVEN datasets.

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 for visual decoding from neurophysiological signals, improving brain-computer interfaces by emulating human visual system processing.

A discrete diffusion model that predicts entire sets of missing knowledge graph triples in one pass, ensuring consistency and achieving state-of-the-art performance.

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

Develop a multimodal foundation model for predicting patient outcomes and improving hospital operations using integrated health records.

Stratagem learns transferable reasoning in language models via trajectory-modulated game self-play, improving performance on mathematical, general reasoning, and code generation benchmarks.

A framework for post-training language models that encourages optimistic exploration and synergizes exploration with exploitation for improved generalization and diversity.

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

This research demonstrates that targeted data augmentation, not architectural limitations, is key to improving LLMs' formal linguistic competence, with code available.

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

Generate adaptable and scalable football tactics using a multi-agent diffusion transformer, grounded in game context and expert validation.

An autonomous web agent framework that tackles dual-level uncertainty in planning and reasoning using adaptive planning and Monte Carlo Tree Search for robust decision-making.

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.

WISV is a wireless-informed semantic verification framework for distributed speculative decoding in 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.

Bounded Ratio Reinforcement Learning (BRRL) provides a theoretically grounded framework for policy optimization, outperforming PPO in stability and performance.

A text-conditioned video-to-music generation model that balances musical fidelity and semantic understanding with fine-grained creator control and faster inference.

An autonomous drone system using YOLO for rapid drowning swimmer detection and localization, significantly reducing rescue response times.

ContraPrompt optimizes LLM prompts by analyzing the reasoning traces of successful and failed attempts, significantly outperforming existing methods on multiple benchmarks.

Agent-World is a self-evolving training arena that synthesizes realistic environments and tasks to co-evolve general-purpose AI agents, outperforming proprietary models on challenging benchmarks.

A framework for analyzing LLM failures on realistic benchmarks using contrastive attribution.

A paradigm to improve VLM-as-a-Judge reliability by balancing informativeness and image-grounded correctness.

An agentic framework that synergizes database probing and rule-based verification to improve text-to-SQL accuracy and efficiency for complex queries.

RAVEN is a framework using LLM agents and RAG to synthesize comprehensive vulnerability analysis reports for memory corruption in code and binaries.

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

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

A training-free control architecture enhances LLM reasoning by intelligently managing memory, improving arithmetic benchmarks significantly.

PLAG enhances tabular anomaly detection by using pseudo-labels to generate localized, feature-level anomalies, achieving state-of-the-art performance and boosting existing unsupervised detectors.

Leveraging Qwen multimodal models with LoRA for improved change visual question answering in remote sensing imagery.

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 using lightweight LLMs, enabling real-time, local inference for sensitive mental health data.

Style-based neural network architectures, including Multi-PatchGAN and Truncated ResNet50 with Gram Matrix and Attention, for real-time weather classification and other appearance-based tasks.

A curriculum framework for AI agents that organizes capability development across scientific, humanities, and social domains, demonstrating improved security and social reasoning.

A time-aware LLM framework that integrates structured EHR encoders via prompt tuning to capture longitudinal patient information and population-level patterns for clinical prediction.

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

A regularization technique that improves next dialogue act prediction in counselling conversations by incorporating empirical dialogue-flow statistics.

LQM is a linguistically motivated framework for evaluating machine translation quality, designed to capture dialect- and culture-specific errors in diglossic languages like Arabic.

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

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 independent noise predictions.

TPS-CalcBench is a diagnostic benchmark and evaluation framework for LLM analytical calculation competence in safety-critical aerospace engineering, including intervention methods.

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

A portable protocol for validating LLM confidence signals, adapted from clinical psychology, can be applied across benchmarks and probe formats.

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.

A physics-consistent dataset and verification-aware reinforcement learning approach to enhance LLM reliability in scientific domains.

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

AQPIM quantizes LLM activations directly in memory to drastically reduce decoding latency and improve speed for Processing-in-Memory architectures.

A co-evolutionary framework that evolves agent architectures and reasoning paths for automated optimization tasks, outperforming existing methods.

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-evolving framework for LLMs to achieve state-of-the-art emotion recognition and consistent expression in conversations through self-play and reinforcement learning.

A benchmark and dataset for evaluating the stylistic robustness of frontier model safety refusals, revealing significant weaknesses in generalization.

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

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

STaD is a framework for generating scaffolded tasks to systematically identify and visualize compositional skill gaps in LLMs.

A benchmark and study of LLM-driven smartphone automation failures, revealing insights into multimodal vs. text-only inputs and common error patterns.

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

Leverage class-specific diffusion models and structural guidance to generate synthetic data for significantly improving military object detection in low-data scenarios.

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 based framework for arbitrary-scale hyperspectral image super-resolution that adaptively reconstructs spatial details while preserving spectral fidelity.

Analyzes distinct jailbreaking methods for open-weight LLMs, revealing divergent behavioral and mechanistic properties despite similar harmfulness.

A diffusion-guided state-space model for multimodal salient object detection that improves boundary accuracy and outperforms existing methods.

A framework to prevent safety alignment erosion in LLMs during continual domain adaptation.

An AI framework enables radiation-free 3D spine reconstruction from MRI for pediatric scoliosis assessment, significantly reducing processing time and improving accuracy.

A logic-based explainable AI framework for 5G intrusion detection that integrates Transformer models with XAI techniques to achieve high accuracy and transparent reasoning.

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

This paper demonstrates that Test-Time Scaling and Supervised Fine-Tuning can bridge the reasoning gap in Vietnamese Small Language Models, outperforming complex agentic workflows for edge deployment.

Identifies diversity collapse in multi-agent LLM systems due to structural coupling, offering insights for designing more creative AI collaborations.

Enabling zero-shot ship instance segmentation in SAR imagery by prompting foundation models with bounding boxes from a SAR-trained detector.

A lightweight inference-time framework for LLMs that uses real-time entropy monitoring to detect and rectify logical failures during long-chain reasoning without compromising fluency.

Revolutionize content moderation by detecting harmful content using internal representations of LLMs for improved safety.

CADMAS-CTX is a framework for contextual capability calibration in multi-agent delegation, improving teamwork by adapting agent capabilities to task context.

A new framework and system for evaluating and fine-tuning LLM agents that eliminates reward hacking and reduces hardware requirements.

A Generalized Synthetic Control Method for baseline estimation in demand response services, outperforming existing methods by treating it as a dynamic counterfactual prediction problem.

KnowledgeBerg is a new benchmark to evaluate LLMs on systematic knowledge coverage and compositional reasoning, revealing significant limitations.

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

A novel prompting technique that maximizes information per token to improve LLM accuracy and reduce hallucinations without additional tokens or latency.

A GenAI platform for policy 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.

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

A deep operator network framework that learns a surrogate for the Riccati solution operator, enabling fast online evaluation of optimal feedbacks for Linear Quadratic Regulator problems.

A data augmentation strategy for training smaller language models to perform topic-controlled educational summarization, improving performance with less data.

A pipeline-adapted reward model that aligns LLM rewards with downstream pipeline execution outcomes for improved consistency in multi-stage LLM applications.

MARC is a novel framework that compresses LLM representations for recommendation systems by controlling modularity, achieving significant online A/B test lift.

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

IceBreaker generates personalized conversation starters for AI agents, overcoming the initial user engagement barrier and demonstrably increasing user activity and click-through rates in production.

A benchmark and memory-augmented method for LLM agents to personalize tool calling by representing user preferences as evolving hypotheses, improving accuracy with minimal token usage.

This paper proposes model-native skills derived from internal representations for more effective LLM intervention and behavior modification.

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

A commit-open protocol uses sparse autoencoder feature traces to detect silent model substitutions in hosted LLMs, outperforming existing methods.

Investigates conditions under which LLMs can outperform human annotators in estimating aggregate subgroup opinions on subjective tasks.

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

Integrating explicit physical feasibility supervision into Vision-Language-Action models improves robot policy reliability and learning efficiency.

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

This work introduces a methodology for creating agile, data-driven AI model documentation to improve model reusability.

LLM agents are trained for spontaneous, reward-free self-evolution by exploring world knowledge, leading to significant performance gains on downstream tasks.

A novel distributed learning framework for diverse and discriminative representations in large-scale scenarios, theoretically proven to maintain desired properties.

RASP-Tuner, a retrieval-augmented soft prompt method for efficient context-aware black-box optimization in non-stationary environments.

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

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

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

An interpretable machine learning framework for stage-aware Parkinson's disease severity classification using multimodal clinical data.

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 quality and visual outputs.

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

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

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

Implicit neural representations offer a coordinate-based framework for continuous environmental field reconstruction from sparse ecological data.

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

Fine-tuning Llama models with LoRA on religious texts to analyze differentiated ethical reasoning patterns.

Proposes new metrics to evaluate and generate narrative explanations for AI, moving beyond feature importance to provide more understandable 'why' behind predictions.

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

A theoretical framework using the Riemann zeta function to predict when additional data will substantially improve performance in biomedical discovery.

A new approach to agentic forecasting using Bayesian updating to estimate future events with linguistic data.

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

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 uses controlled routing tasks to reveal how prompt-based interventions alter LLM behavior and identify where relevant state is represented.

Proposes party autonomy as a solution for determining applicable law in cross-border data transfer disputes, aligning non-contractual obligations with contractual choices.

Investigates secret leakage risks in code LLMs, identifying a 'gibberish bias' in BPE tokenization as a root cause for memorization.

This paper explores the vulnerability of ranking systems to strategic data manipulation, proposing an attack method to identify high-impact perturbations and highlighting the need for more robust aggregation methods.

An evolutionary tree-based generator for creating psychometric contexts to assess creativity.

A domain-specific language and framework for provable coordination of LLM agents, ensuring deadlock-free communication.

Demonstrating that randomly initialized networks can learn representations through peer-to-peer consensus without complex mechanisms.

A runtime security architecture for LLM agents that improves robustness against prompt-injection attacks.

This study uses LLM agents in simulated markets to reveal behavioral finance patterns and demonstrates how prompt interventions can control market bubble formation.

A framework for generative music models that uses a latent-space Fourier transform to provide frequency-domain controls for timescale-based manipulation and blending.

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

A community-led framework for integrating AI into wildfire risk assessment with a focus on literacy and explainability.

A novel framework for large language models that applies reinforcement learning directly to continuous latent representations to improve reasoning diversity and performance.

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

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

A framework for auditing AI-generated code to detect failure-untruthful patterns, crucial for safety-critical systems.

This research identifies key factors like stochasticity, ambiguity, and behavioral variability that cause unreliability in computer-use AI agents.

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

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

This paper investigates methods to improve policy gradient reinforcement learning by addressing discontinuities in differentiable simulators, proposing new estimators for better performance.

A novel logic-to-topology encoding framework to bridge formal logic, topology, and neural processing for mechanistic interpretability in neuro-symbolic AI.

A calibrated reinforcement learning approach for multi-attempt chain-of-thought reasoning that optimizes verification success rates.

This survey reviews classical multi-agent systems and explores their evolution towards foundation model-enabled futures, outlining challenges and opportunities.

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

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

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

This paper explores the theoretical foundations of optimization algorithms by drawing parallels to natural physics and non-Newtonian dynamics.

A conceptual framework for understanding the fragility of human-AI collaboration, identifying factors beyond model capability that impact stable interaction.

This paper theoretically analyzes the depth-cache tradeoffs in KV-compressed Transformers, focusing on memory bottlenecks during inference.

Analysis of polarization and integration in global AI research over three decades, highlighting US and China's diverging influence.

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

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

A minimal neural cellular automaton spontaneously develops a structured internal representation for syntactic processing, mimicking CKY algorithm principles.