Build Loop

A benchmark for frontier coding agents to autonomously implement ML pipelines, demonstrating significant progress in AI self-improvement capabilities.

A new dataset and fine-tuned models for 3D airway bifurcation detection in CT scans, addressing a critical gap for respiratory disease analysis.

OxyGent is an open-source framework for building modular, observable, and evolvable multi-agent systems with a Lego-like assembly paradigm.

A new dataset and pipeline for training Retrieval-Augmented Generation models to prioritize retrieved context over internal knowledge.

Exploratory Sampling (ESamp) is a decoding approach for LLMs that boosts semantic diversity and reasoning performance by using a lightweight distiller to predict novelty signals, with less than 5% overhead.

Automated platform for generating animated video stories from data.

An AI-assisted scientific computing package for calculating cosmic microwave background spectral distortions, offering a novel approach to scientific software development.

An open-source pipeline for learning illumination control in diffusion models, enabling precise image adjustments via natural language instructions.

A framework for distilling large audio foundation models into smaller, deployable versions using only output embeddings, reducing size by up to 61x while retaining 96% performance.

A specialized Korean legal LLM built with domain experts and use-case-driven data curation for practical legal tasks.

A topology-faithful dimensionality reduction method that preserves global structure better than UMAP at scale, with a new benchmark for evaluation.

A refined workflow and dataset (GAIA-v2-LILT) for creating valid multilingual agent benchmarks that align functionally and culturally, improving agent performance by up to 32.7% over machine-translated versions.

A plug-and-play method to reduce hallucinations in vision-language models by intervening during the prefill stage, improving initial representations.

A deployed semantic search system for health systems that retrieves clinical information from millions of notes with sub-second latency and significant cost savings.

TSN-Affinity is a continual offline reinforcement learning method that uses similarity-driven parameter reuse to prevent catastrophic forgetting and improve multi-task performance.

A zero-trust semantic gateway for secure enterprise AI-native systems that validates autonomous agents as stochastic state-transition systems.

Distill complex multi-agent debate into a single LLM for improved reasoning and controllable behavior, reducing token usage by up to 93%.

QFlash enables end-to-end integer-only FlashAttention for Vision Transformers, significantly speeding up computation and reducing energy consumption.

AHASD is a novel asynchronous heterogeneous architecture for mobile LLM speculative decoding, significantly improving throughput and energy efficiency.

Cooperate to Compete (C2C) is a novel multi-agent environment and dataset for studying strategic negotiation and coordination in LM-based agents, with implications for real-world deployments.

Agentic Architect optimizes microarchitecture design using AI-driven evolutionary models.

A challenging benchmark for multimodal, multi-entity, multi-hop visual question answering to evaluate large language models' reasoning capabilities.

Enhancing federated learning of LLMs on edge devices with Fisher-Guided Token Quantization for significant communication efficiency and speedup.

A recursive multi-agent framework that enhances reasoning and efficiency in complex tasks through iterative collaboration and latent state transfer.

A tailored recommendation engine using action-aware generative modeling for short video platforms to enhance user retention.

Canonical knowledge distillation significantly improves semantic segmentation performance, achieving state-of-the-art results with smaller models by matching compute budgets.

LLM-ReSum is a self-reflective summarization framework that integrates LLM-based evaluation and generation in a feedback loop, improving factual accuracy by up to 33% and coverage by 39%.

BenchGuard is an automated auditing framework that uses LLMs to find critical flaws in agent benchmarks, making AI development more reliable and cost-effective.

Open-source, highly efficient multilingual Mixture-of-Experts language models with a strong performance-to-compute ratio.

A multi-agent framework that decomposes instructed code editing into specialized roles to improve reliability and reduce unintended changes.

A hierarchical multi-agent framework that optimizes video storytelling for coherent narratives and personalized advertising.

A benchmark and metric for evaluating LLM-generated REST API tests based on natural language requirements, improving functional validation.

Luminol-AIDetect efficiently identifies AI-generated text by leveraging unique structural signals.

Improves generative model alignment with complex human preferences by using a semi-supervised approach to handle noisy preference data.

A multi-source benchmark and evaluation pipeline for assessing the structured output quality of large language models across text, image, and audio inputs.

Scales visual preference optimization for generative models by introducing a novel adaptive algorithm and a massive, high-quality preference dataset.

Develop a tool for creating interactive courseware with a generative user interface.

Generate synthetic training data for custom LLM guardrails using a debate framework, eliminating the need for extensive human annotation.

A framework that uses interpretability insights to select data that maximally activates internal task features, leading to highly efficient LLM fine-tuning.

A 3D hierarchical music generation model that creates controllable orchestral pieces with advanced harmony and reduced dissonance.

Physics-informed test-time adaptation for robust human activity recognition on mobile devices, improving accuracy and stability.

A benchmark and fine-tuned LLM for automatically evaluating K-12 science instructional materials.

A reinforcement learning framework for cold-start LLM personalization that jointly reasons over user writing styles and topic preferences.

This research presents a novel white-box jailbreaking technique for LLMs that directly optimizes prompt embeddings, outperforming existing methods by preserving prompt semantics while achieving higher success rates.

Lightweight and fast prompt injection detection for screenshot-based web agents, outperforming larger models.

An automated framework for synthesizing diverse terminal tasks using skill graphs to train more capable command-line agents.

A benchmark for reliable semiconductor bandgap prediction, bridging the gap between computational models and experimental measurements.

Analyzes weak-to-strong AI alignment failures using a bias-variance perspective, identifying strong-model variance as a key predictor of deception.

Generative AI integrated with extreme value theory for tail-aware channel estimation in URLLC networks.

A friction-aware reinforcement learning framework for zero-shot sim-to-real transfer in humanoid ballbots, enabling expressive maneuvers via a generalized iOS ecosystem.

A tool to evaluate code plagiarism detection metrics, offering a path to more reliable academic integrity checks in software engineering.

An LLM agent framework that automates end-to-end 3D cutscene generation by integrating with game engines and orchestrating specialist sub-agents.

A neurocognitive governance framework for autonomous AI agents that embeds self-governance principles into their reasoning, achieving 95% compliance accuracy in a retail supply chain workflow.

The first dataset of GPT-Image-2 generated images from Twitter, analyzing their characteristics and the challenges of content provenance.

An ultra-low-power FPGA-based CNN for on-device cardiac feature extraction in smart health sensors, enabling autonomous health monitoring for astronauts.

Analyzing LLM reasoning steps to uncover and categorize hidden mental health stigma, offering a more nuanced evaluation than traditional methods.

Libra-VLA is a dual-system VLA architecture for robotics that decouples coarse-to-fine planning and refinement for scalable, robust, and responsive open-world manipulation.

R3-SQL is a Text-to-SQL framework that improves query ranking consistency and recall through unified reward and agentic resampling, achieving state-of-the-art performance.

PSI-Bench provides clinically grounded and interpretable evaluation for depression patient simulators, identifying key limitations and guiding future development.

An agentic workflow using LLMs and structured queries to automate the initial investigation of security alerts, improving accuracy and reducing analyst workload.

A speaker-adaptive network for emotion recognition in conversations that calibrates input, gates modality trust, and regularizes speaker features for improved accuracy.

An empirical study of uncertainty estimation for audio-aware large language models to improve reliability and detect hallucinations.

A multimodal framework that uses code and comments to significantly improve software vulnerability detection accuracy.

DRAGON is a new benchmark and dataset for evaluating evidence-grounded visual reasoning in diagrams, addressing limitations of current vision-language models.

A label-free reinforcement learning framework for LLMs that decouples answer proposal from reward verification, improving mathematical reasoning and code generation.

A decoupled reinforcement learning framework for reasoning-driven image editing that optimizes planning independently from generation.

A sample-efficient neuro-symbolic PPO that uses logical policy specifications to guide learning in sparse-reward reinforcement learning domains.

SIEVES improves multimodal LLM coverage on out-of-distribution tasks by scoring the quality of visual evidence used in predictions.

A production-ready inference architecture for scalable and cost-effective deployment of complex, multi-model AI systems.

A continual learning framework for brain lesion segmentation that integrates visual features with structured concepts for improved adaptation and interpretability.

Enables parameter-efficient multi-anchor word representations in large language models through novel projection and encoding techniques.

A real-time adaptive traffic signal system that uses YOLOv12 and ByteTrack to extend crossing times for vulnerable road users, improving safety by 71.4%.

A tool to measure model sensitivity to data errors, enabling prioritized data cleaning for improved machine learning performance.

CORAL is an adaptive retrieval methodology for multilingual RAG that iteratively refines retrieval space and query for culturally-aligned answers, improving accuracy by up to 3.58%p on low-resource languages.

A watermark removal detection method that identifies hidden statistical artifacts, making current removal techniques forensically detectable.

A large-scale Dutch medical language corpus released on Hugging Face to enable NLP development in the Dutch medical domain.

A benchmark demonstrating that providing explicit semantic context significantly improves LLM accuracy for data analytics, reducing hallucinations.

A reinforcement learning framework that learns expert rewards directly from expert chip layouts to achieve expert-level placement.

An orchestration architecture for LLM agents that improves long-horizon knowledge synthesis by managing knowledge states and evidence chains.

Nemotron 3 Nano Omni is an open-source multimodal AI model with native audio support, offering improved accuracy and efficiency for real-world applications like document understanding and agentic computer use.

Evaluating 15 frontier AI chatbots on psychiatric triage, this study reveals high accuracy for emergencies but significant over-triage for lower-risk presentations.

A recursive sparse mixture-of-experts framework for multimodal diffusion models that enhances structured reasoning in text-to-image generation.

A novel framework that decouples retrieval and reranking for multi-modal knowledge graph completion, improving accuracy.

A framework to analyze and mitigate y-axis bias in multimodal language models for chart-to-table translation, improving performance and fairness.

Optimize operations research problems with AI via decentralized debate and memory-enhanced agents.

TrialCalibre is a conceptualized multi-agent system to automate and scale causal inference for RCT benchmarking and observational trial calibration.

A multi-task EEG analysis framework using low-rank adaptation to efficiently adapt pre-trained models to multiple downstream tasks.

GraphPL, a graph neural network approach for unsupervised modality imputation in patchwork learning settings, demonstrating state-of-the-art performance on benchmark and real-world datasets.

A language-agnostic system to detect cross-lingual LLM jailbreaks using semantic similarity without retraining.

Improving small language model reasoning under compute and token constraints through stepwise guidance.

Cooperative path planning for ground and aerial vehicles to navigate uncertain road networks, useful for disaster response.

Demonstrates that a small transformer can be trained to execute programs in a universal programming language, acting as a universal computer.

This research demonstrates that unstructured pruning can augment test-time reasoning performance in LLMs, challenging the notion that pruning always degrades capabilities.

VAE-Inf is a novel framework for imbalanced classification that combines variational autoencoders with hypothesis testing for statistically interpretable and robust predictions.

This paper categorizes imperfect rewards in policy gradient optimization, showing that errors can be benign or beneficial, and proposes new evaluation metrics for reward models.

Generative UI can create adaptive interfaces for e-commerce platforms, bridging accessibility gaps for users with disabilities.

A framework that uses structured specifications and optimization to design networked systems, outperforming LLMs in constraint satisfaction.

An LLM-powered system for accurate recipe nutrient estimation, balancing precision with computational efficiency for dietary monitoring.

A novel graph-guided loss function that improves fine-tuning of language models by incorporating global semantic structure for better classification accuracy.

Introduces HotComment, a multimodal benchmark and StyleCmt model for evaluating online comment popularity by considering content quality, trends, and user behavior.

A novel document sampling method for unsupervised domain adaptation in information retrieval that prioritizes model uncertainty to improve generalization.

An investigation into how linguistic biases in different English dialects affect LLM recommendations for restaurants and products.

A hybrid framework combines machine learning with non-monotonic reasoning to improve context-aware phishing website classification and allows for efficient knowledge updates.

Evaluating LLMs' understanding of embodied cognition and cultural variation using cross-linguistic demonstratives.

A genetic programming algorithm for multi-task reinforcement learning in continuous control environments with interpretable decision flows.

This research investigates positional bias in small LLMs as a failure mode for detecting deliberate underperformance, suggesting positional distribution shifts as a more effective signature.

Improves zero-shot reinforcement learning by sampling task vectors directly from offline datasets, enhancing generalization.

A novel loss function continuum for reasoning models that mitigates cold-start issues in fine-tuning by balancing exploitation and density estimation.

Reveals hidden linear-centroid coupling in neural networks by analyzing gradient directions, impacting feature formation and grokking.

A systematic empirical study investigating the scaling behavior and geometric conditions for in-context learning in Transformers.

An AI-driven framework for online task offloading in mobile edge computing networks that optimizes computing and energy resources.

A novel system combining AI and cryptography for scalable, secure biometric authentication without auxiliary identifiers, protecting against data breaches.

A semi-Markov reinforcement learning approach for city-scale EV ride-hailing that guarantees feasibility of actions and optimizes net profit.

SkillGuard-Robust is a novel system for auditing untrusted agent skills, improving security review consistency and robustness against semantic-preserving rewrites.

A transformer-based meta-agent that learns to model and exploit opponents in imperfect-information games by adapting its policy towards best-response exploitation.

Investigating the use of LLMs to create dynamic, emotionally receptive work environments by adjusting physical and virtual settings.

A methodology for creating digital twins to evaluate the security of autonomous platforms against various threats.

An automated adversarial collaboration framework uses LLMs and program synthesis to adjudicate between competing scientific theories.

Demonstrates how large language models can be manipulated to spread misinformation by prioritizing fringe scientific material, posing risks to public understanding.

A dependency-driven prompt pipeline for generating coherent and scalable RPG content using LLMs.

A framework for transforming partially specified human intent into inspectable artifacts for open-world AI agents by formalizing closure gaps and defining delegation envelopes.

A Q-learning framework for safe reinforcement learning that prevents unsafe state visitation during training.

Investigates the persistence of gradient alignment in multi-step settings for subliminal learning, showing current mitigation methods may not reliably suppress trait acquisition.

Ensuring LLM formalizations are faithful through roundtrip verification and targeted repair.

A faceted model for transparent attribution of AI-assisted text production, detailing the form, generation, evaluation, intent, control, and traceability of AI interventions.

A medoid prototype alignment framework for detecting unknown attacks in industrial control systems across different plants.

A novel approach to test and attribute failures in embodied agents by focusing on specific capabilities.

A theoretical framework for bandit problems with smooth payoffs on graphs, applicable to content-based recommendation.

A protocol for secure delegation of user secrets in agentic systems to enable authorized operations without exposing reusable authority.

An empirical study exploring how university students integrate AI tools into diverse academic writing workflows, identifying three distinct usage configurations.

Explains how overparameterization in neural networks improves optimization through weight-space symmetries.

A framework for evaluating factual correctness in procedural video captions, highlighting limitations of current multimodal models.

A Bayesian framework to embed linear relationships into machine learning for improved uncertainty estimates and adherence to physical constraints.

GPT-based tools show promise for assisting in spreadsheet model building but remain unreliable for professional use due to inconsistencies and workflow challenges.

This paper introduces a new training methodology for LLMs that aligns training objectives with test-time inference strategies to improve performance.

ValueAlpha is a stress-testing protocol for LLM-judged investment rationales, ensuring claims are stable and agreed upon before observable returns.

MultiHedge is a hybrid architecture using LLMs and retrieval to improve robustness and stability in modular decision-making systems.

This paper analyzes agentic AI failures in astrophysical workflows, highlighting silent incorrect computations as a critical risk and releasing an evaluation framework.

A TEE-based architecture for auditable AI grant evaluation without exposing the model, focusing on remote attestation and prompt injection risks.

Investigates the functional perspective of layer redundancy in LLMs, showing calibration objectives influence pruning more than search algorithms.

An exploratory study on the barriers and enablers of online instruction in hospitality education, highlighting the need for AI training and support.

Develops an effective field theory to explain discrepancies in Kohn-Sham eigenvalues and ARPES measurements for metals, offering a computationally inexpensive correction.

This research identifies conditional misalignment in language models, where common interventions can hide emergent misbehavior that reappears when prompts resemble training data.

Proposes geometric algebra as a superior mathematical foundation for natural language semantics, offering enhanced compositionality and interpretability over linear algebra.

This paper proposes a framework for accountable knowledge distillation, focusing on preserving critical teacher model capabilities beyond simple task scores to ensure reliable and safe student models.

A per-task leverage ratio framework for human-agent collaboration, analyzing information flow and task novelty to optimize efficiency.

This paper theoretically distinguishes three models of human feedback in RLHF to guide annotation strategies and identify failure modes.

Develops algorithms for designing strategic audits to mitigate fraud in resource allocation domains by modeling audit policy design as a principal-agent game.

Develop a systematic recipe for translating ReLU approximation results to softmax attention mechanisms in transformers, providing new analytical tools.

This paper argues for a co-design agenda between AI tools and MBSE models to enable AI participation beyond simple prompt-based interaction.

Overview of Internet of Everything paradigms, enablers, and future directions in the 6G era.

A comprehensive framework for inferring learner cognitive and affective states from observable nonverbal cues, based on a systematic review of existing research.

A framework for learning language model policies that manage epistemic and normative risk through explicit intervention actions.

This paper explores value-sensitive AI systems for prayer, focusing on preserving user agency and authenticity in spiritual contexts.

Theoretical introduction to the verification of neural networks, covering feed-forward, recurrent, attention mechanisms, and transformers.

Establishes expressiveness relationships between different types of Recurrent Graph Neural Networks and introduces a coordination protocol for asynchronous halting.