Build Loop

A challenging benchmark for multimodal, multi-entity, multi-hop visual question answering that reveals significant limitations in current large language models.

A new dataset and pipeline to train RAG models to prioritize retrieved context over internal knowledge, improving faithfulness.

Automate data-driven video story generation with a multi-agent framework and a new benchmark for evaluating animated visualizations and narration.

An open-source AI framework that uses LLMs and simulation to explore and optimize computer architecture designs, outperforming state-of-the-art.

Develop an AI platform for strategic coordination and multiplayer game scenarios to enhance cooperation in competitive environments.

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

A zero-trust security gateway for AI-native enterprise systems that formally validates and audits autonomous agents using semantic fuzzing.

AHASD is a mobile NPU-PIM heterogeneous architecture for speculative decoding that uses task-level decoupling and adaptive control to improve LLM inference throughput and energy efficiency.

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 health-system-scale semantic search system for clinical notes that reduces chart abstraction time by up to 89% with sub-second latency and low operational costs.

QFlash enables end-to-end integer quantization for Vision Transformer attention, achieving significant speedups and energy reduction without accuracy loss.

AutoResearchBench revolutionizes AI-driven scientific literature discovery with a challenging new benchmark.

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

A generative sequence network that refines user actions into temporal sequences for nuanced short video recommendations, achieving significant improvements in user engagement and retention.

A recursive multi-agent framework that enhances collaboration and efficiency through iterative refinement and latent state transfer, outperforming baselines in reasoning and speed.

A dataset of self-reported AI-generated images from the first week of GPT-Image-2 deployment, revealing challenges in provenance verification on social media.

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%.

Open-source, highly sparse multilingual Mixture-of-Experts language models with efficient upcycling and best-in-class performance-to-compute ratio.

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

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

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

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

A fast, zero-shot detector for machine-generated text that uses perplexity under text shuffling to identify structural fragility.

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

A framework that guides data selection for LLM fine-tuning by identifying and activating internal task features, leading to significant data efficiency gains.

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

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

A Korean legal-domain LLM specialized through a use-case-driven dataset construction and optimized training pipeline, collaborating with legal professionals for precision and reliability.

PSI-Bench is an interpretable evaluation framework for depression patient simulators, providing clinically grounded diagnostics to guide future development and improve realism.

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 VLMs in grounding their predictions.

SymphonyGen is a 3D hierarchical framework for generating cinematic orchestral music with controllable harmony skeletons and improved musicality.

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

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

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 benchmark for semiconductor bandgap prediction that bridges the gap between computational and experimental data, revealing generalization limitations.

GraphPL, a graph neural network approach for robust modality imputation in patchwork learning, achieving state-of-the-art performance on distributed electronic health record data.

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.

R3-SQL is a Text-to-SQL framework that ranks functionally equivalent queries consistently and uses agentic resampling to improve candidate recall, achieving state-of-the-art execution accuracy.

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

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

A Continual Learning framework for brain lesion segmentation that integrates visual features with structured concepts to simulate clinical reasoning and guide model growth.

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

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

Neuro-symbolic PPO that uses logical policy specifications to accelerate learning and improve performance in sparse-reward reinforcement learning tasks.

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.

This research introduces a new metric for watermark removal detection, showing current methods leave forensic artifacts and no existing method balances removal success, quality, and detectability.

SIEVES improves multimodal LLM coverage on out-of-distribution tasks by learning to score the quality of visual evidence for selective prediction.

This system uses real-time AI to detect and track vulnerable road users, extending traffic signal timing to prevent pedestrians from being stranded.

Optimized communication-efficient federated learning solution to fine-tune language models on edge devices.

A benchmark demonstrating that providing explicit semantic context significantly improves LLM accuracy and reduces hallucinations in data analytics.

A production-tested inference architecture for enterprise compound AI systems, reducing latency and costs while enabling rapid iteration.

An architecture for long-horizon knowledge synthesis that orchestrates LLM agents with explicit state management and checkpointing for reliable task completion.

The first large-scale Dutch medical language corpus, comprising 35 billion tokens, is now available on Hugging Face for pre-training and downstream NLP tasks.

Generative UI can create adaptive interfaces for user-generated content platforms, bridging accessibility gaps for visually impaired and older users.

Develop a platform for optimizing complex models using memory-enhanced decentralized debate among AI agents.

SnapGuard is a lightweight method for detecting prompt injection attacks in screenshot-based web agents by analyzing visual and textual signals.

Evaluating 15 frontier AI chatbots for emergency psychiatric triage, finding high accuracy for emergencies but over-triage for lower-risk cases.

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

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

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

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

An open-source pipeline for controlling illumination in diffusion models using natural language instructions, outperforming existing models.

TrialCalibre is a multi-agent system designed to automate and scale the process of benchmarking and calibrating observational trials against RCTs for credible real-world evidence studies.

This research evaluates existing code evaluation metrics for their effectiveness in detecting source code plagiarism, suggesting potential for improved automated plagiarism detection tools.

A framework for synthesizing diverse terminal tasks using skill graphs to improve agent training for command-line execution.

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

Investigating linguistic biases in LLM recommendations across different English and Hindi-English dialects.

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

An empirical study of uncertainty estimation for audio-aware large language models to improve reliability in audio understanding tasks.

VAE-Inf is a two-stage framework for imbalanced classification that uses a variational autoencoder on majority data to build a reference distribution, then fine-tunes with minority data for statistically interpretable hypothesis testing.

A cooperative path planning system for ground and aerial vehicles to navigate uncertain road networks, demonstrated on 100 urban road networks.

Frontier coding agents can now autonomously implement complex ML pipelines, demonstrating a potential early warning signal for recursive self-improvement in AI research.

A refined workflow and dataset for creating valid multilingual agent benchmarks that go beyond simple translation, improving agent performance by up to 32.7% over machine-translated versions.

A novel metric and toolset to identify and prioritize data errors that most impact machine learning model performance, enabling targeted data cleaning.

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

BenchGuard is an automated auditing framework that uses LLMs to find critical flaws in agent benchmarks, significantly reducing costs and improving reliability.

This paper categorizes imperfect rewards in policy gradient methods, showing that some errors can be beneficial for language model training.

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

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

Canonical knowledge distillation achieves state-of-the-art semantic segmentation performance with significantly less compute than current complex methods.

Investigating the use of LLMs to dynamically adjust work environments for enhanced focus, well-being, and engagement in hybrid settings.

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

SkillGuard-Robust is a novel system for auditing untrusted agent skills, achieving high accuracy in detecting malicious intent and ensuring robustness against semantic rewrites.

A graph-guided loss function that improves fine-tuning of language models by incorporating global semantic structure for more discriminative embeddings.

Improves zero-shot offline reinforcement learning by extracting task vectors from the dataset for more principled policy training.

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

Distill multi-agent debate into a single LLM for improved reasoning and controllable harmful behavior, achieving performance with significantly fewer tokens.

Scales visual preference optimization for generative models by introducing a novel dataset and a robust training objective that adapts to data quality.

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

A new dataset and baseline models for 3D airway bifurcation detection in CT scans, enabling specialized tools for respiratory disease analysis.

Exploratory Sampling (ESamp) is a decoding approach for LLMs that uses a lightweight distiller to encourage semantic diversity during generation, boosting reasoning and creative writing performance with minimal overhead.

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.

This paper compares traditional and LLM-based methods for estimating recipe nutrients, highlighting a trade-off between accuracy and computational cost.

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

A novel loss function continuum that mitigates cold-start stalling in reasoning models by dynamically adjusting supervision commitment.

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

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

Enhancing URLLC in 5G+ networks by integrating extreme value theory with generative AI for tail-aware channel estimation.

This research analyzes LLM reasoning steps to uncover and categorize hidden mental health stigma, offering a more robust evaluation than traditional methods.

A friction-aware reinforcement learning system for humanoid ballbot hardware, achieving zero-shot Sim2Real transfer.

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

This research introduces a novel white-box jailbreaking technique for LLMs that directly optimizes prompt embeddings, preserving semantic content and outperforming existing attacks.

An AI-driven framework for offloading tasks in mobile edge computing networks to optimize energy and computing resources.

An empirical investigation into the in-context learning capabilities of transformers, focusing on scaling behavior and benign overfitting.

A semi-Markov reinforcement learning approach for city-scale EV ride-hailing that guarantees feasibility of dispatch, repositioning, and charging decisions.

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 Q-learning framework for reinforcement learning that eliminates unsafe state visitation during training by leveraging a behavior policy supported on a safe set.

A transformer-based meta-agent that learns to model and exploit opponents in imperfect-information games by adapting its strategy based on opponent behavior.

A threat-oriented digital twinning methodology for evaluating the cybersecurity of autonomous platforms, adaptable for UAV and space systems.

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

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.

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

Nemotron 3 Nano Omni is an open multimodal AI model with native audio support, offering improved accuracy and efficiency for real-world applications.

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

Analyzes how overparameterization in neural networks introduces symmetries that improve optimization and accelerate convergence.

A new method for attributing failures in vision-language navigation agents to specific capabilities, improving interpretability and guidance for agent development.

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

A theoretical framework for bandit problems with payoffs smooth on a graph, applicable to online learning and content-based recommendation.

A framework for evaluating factual correctness in procedural video captions by separating conceptual and contextual facts.

A novel method for unsupervised domain adaptation in information retrieval that improves document sampling by considering model uncertainty.

GPT-based tools show promise for assisting in spreadsheet model building but remain unreliable for professional use due to inconsistency and reproducibility issues.

This paper evaluates agentic AI systems in astrophysical workflows, highlighting silent failures where systems produce plausible but incorrect results, and releases an evaluation framework.

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 than below-chance accuracy.

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

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

This research identifies conditional misalignment in language models, where common interventions can mask emergent misbehavior that reappears in specific contexts.

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

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 teacher capabilities beyond simple task scores to ensure reliable student models.

A novel system for scalable, secure biometric authentication against cloud databases, combining AI with cryptography to prevent data breaches.

This paper theoretically distinguishes three models of human feedback for large language models to improve annotation strategies.

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

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

An effective field theory approach to derive quasiparticle band narrowing from frozen core dynamics in metals, potentially applicable to AI-driven scientific discovery.

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

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

Current AI tools interact with MBSE models as prompt engines, not knowledge bases, necessitating a co-design of models and methodologies for true AI participation.

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 risk through explicit intervention actions, going beyond surface-level alignment.

Investigates the persistence of gradient alignment in multi-step learning settings and its causal contribution to unintended trait acquisition in neural networks.

This paper explores value-sensitive AI systems for prayer, arguing for designs that preserve user agency and authenticity by leveraging AI's inexplicability or non-use.

This paper theoretically analyzes the expressiveness of different Recurrent Graph Neural Network models and establishes relationships between them.

A Bayesian framework to embed linear relationships into machine learning for improved uncertainty estimates and constraint satisfaction.

A hybrid architecture using LLMs and retrieval to improve robustness and stability in modular decision systems.

A protocol for agents to use user secrets for operations without exposing reusable authority.

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

Theoretical introduction to the verification of neural networks, covering various architectures and verification techniques.

An exploratory study identifies barriers and enablers for online instruction in hospitality education, highlighting the need for improved pedagogical training and AI integration support.

Investigates the functional perspective of layer redundancy in LLMs, suggesting calibration objectives are more influential than search algorithms for depth pruning.

A per-task leverage ratio framework is proposed to quantify human-agent collaboration efficiency, decomposing information flow and identifying irreducible task novelty.

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

Develops a theoretical framework for translating ReLU approximation results to softmax attention mechanisms in transformers.