Build Loop

A system for constructing debiased multimodal preference data and iteratively training reward models, achieving state-of-the-art performance on multimodal alignment benchmarks.

SCURank enhances summarization by ranking candidate summaries based on information content richness and semantic importance, outperforming traditional methods.

SafetyALFRED evaluates multimodal LLMs for embodied agents, revealing a critical gap between hazard recognition and active risk mitigation in real-world scenarios.

A controllable framework for generating adverse weather videos for autonomous driving, improving perception robustness with high visual quality and annotation reusability.

A multi-agent framework that learns and reuses optimization strategies for generative engines to improve answer quality and citation accuracy.

VLA Foundry provides a unified framework for training and evaluating vision-language-action models in robotics.

A method to accelerate reinforcement learning by efficiently filtering action candidates during the denoising process, reducing computational cost.

An agentic framework and benchmark for generating executable visual workflows from natural language, addressing costly manual development.

A parameter-efficient fine-tuning framework that uses semantic awareness and adaptive scaling to improve multi-task learning for Large Language Models.

A new benchmark and evaluation framework for assessing LLM performance on Indian financial regulatory text, with publicly available code and dataset.

This research evaluates whether advanced LLMs 'game' formalization by generating proofs that are valid but not faithful to the original logical problem, with code and data available for reproducible analysis.

This paper introduces a four-axis decision alignment framework for enterprise AI agents, enabling granular evaluation of factual precision, reasoning coherence, compliance, and abstention, with a transferable methodology for regulated domains.

GRASPrune: A post-training structured pruning framework for LLMs that significantly reduces parameters and inference costs while maintaining performance.

EgoSelf transforms personal long-term interactions into predictive assistants for AR applications.

A reasoning-aware framework for detecting AI-generated content with interpretable explanations and state-of-the-art accuracy.

A novel test set and methodology to expose implicit local and global biases in multilingual LLMs, with code available.

A multi-agent review methodology that uses adversarial agents to discover LLM-assisted software defects with high precision.

A training-free, plug-and-play framework for pruning vision-language models in autonomous driving, achieving state-of-the-art performance with significant compression.

A visuomotor policy for robotic manipulation that learns when and what to recall from memory, significantly improving performance on non-Markovian tasks.

A unified audio front-end LLM that handles multiple speech tasks for seamless, full-duplex conversational AI systems.

OLLM enhances LLM controllability and robustness by replacing single next-token prediction with a set of learned options, enabling more efficient and aligned generation.

This paper introduces a novel probabilistic Gaussian model and adaptive rectification technique for multi-modal test-time adaptation, achieving state-of-the-art performance.

This research provides a comprehensive benchmark and reproducible framework for evaluating LLMs on core social media analytics tasks, with code and data released.

Introduces a novel evaluation framework and benchmark for assessing LLM agent reasoning in financial time-series analysis, demonstrating near-perfect tool-use accuracy.

A federated LLM adaptation framework that uses a proxy SLM and heterogeneity-aware fusion to achieve secure, high-performance fine-tuning without compromising client privacy or LLM IP.

ProjLens is an interpretability framework that demystifies backdoor vulnerabilities in multimodal LLMs by analyzing the role of projector layers.

LePREC is a neuro-symbolic framework that improves legal issue identification by combining LLM-generated question-answer pairs with structured statistical reasoning for relevance assessment.

A component-wise evaluation framework for medical QA systems that reveals health equity implications and performance failures, with code available.

A framework that provides coding agents with instant access to research repositories and technical documentation for real-time, context-aware operation in specialized scientific domains.

A new benchmark for evaluating AI agents on cross-application workflow orchestration via REST APIs, revealing current frontier models score below 10% on realistic business tasks.

A systematic analysis of verbal tics in frontier LLMs, revealing significant inter-model variation and highlighting the 'alignment tax' on authentic human-AI interaction.

A principled framework grounded in game theory for attributing rewards in multi-turn dialogues, enabling language agents to learn social intelligence and achieve state-of-the-art performance.

A framework that unifies task and motion adaptation for human-robot teams, improving efficiency and user experience.

Develops a novel autoregressive model for real-time target speaker extraction that achieves stable performance and superior intelligibility compared to offline baselines.

RoboWM-Bench is a new benchmark for evaluating the physical plausibility and robotic executability of video world models in manipulation tasks.

An open-source benchmark for evaluating LLM agents in realistic cybersecurity Capture The Flag challenges, with a novel partial-credit scoring system.

A Mamba-based multi-agent policy optimization framework for cooperative underwater robot pursuit, improving success rate and efficiency.

A comprehensive benchmark for detecting hallucinations in large audio-language models across speech, sound, and music.

CoDA enables LLMs to effectively transfer knowledge across domains by aligning latent reasoning representations, significantly improving performance in expertise-scarce areas.

Leveraging Low-Rank Adaptation (LoRA) to improve off-policy reinforcement learning critics by constraining updates to a low-dimensional subspace, showing consistent gains in critic loss and policy performance.

A lightweight binarization framework for deploying large language models in resource-constrained environments with state-of-the-art accuracy.

This research introduces personalized benchmarking for LLMs, demonstrating that individual user preferences diverge significantly from aggregate rankings and can be predicted using topic and style features.

PLaMo 2.1-VL: A lightweight, Japanese-capable Vision Language Model for autonomous devices, excelling in VQA and Visual Grounding for industrial applications.

A novel depth pruning method for LLMs that significantly improves efficiency and performance by jointly considering layer similarity and transformation difference.

A novel framework to mitigate cognitive bias in LLM agents by enforcing perspective-invariant reasoning through dialectical alignment, improving fault resolution in ambiguous scenarios.

A multi-party privacy-preserving entity alignment protocol for federated learning that hides intersection membership and supports noisy identifiers.

SAHM is a new benchmark and dataset for Arabic financial and Shari'ah-compliant reasoning, with an instruction-tuned model to improve LLM performance in this domain.

An unsupervised industrial defect detection system using diffusion models for data generation and an asymmetric student-teacher network for precise localization.

LightEdit enables scalable and cost-effective lifelong knowledge editing for LLMs by selectively suppressing outdated information and incorporating new knowledge.

DW-Bench, a new benchmark for evaluating LLMs on data warehouse graph topology reasoning, integrating foreign-key and data-lineage edges, showing tool-augmented methods significantly outperform static approaches.

CoCo-SAM3 enhances open-vocabulary semantic segmentation by resolving concept conflicts between prompts for more stable and accurate multi-class inference.

A modular reinforcement learning framework enables bipedal soccer robots to adaptively control multiple tasks like walking and kicking, with rapid fall recovery.

A unified framework for transferring human motion policies to humanoids by learning a shared latent physical language.

Nexusformer enables stable and inheritable transformer scaling through nonlinear attention expansion, reducing compute and improving performance.

A unified ASR framework with consistency regularization reduces the gap between offline and streaming performance, offering cost-effective and scalable speech recognition.

A neuro-symbolic framework that restructures statement autoformalization into a modular pipeline, improving accuracy and error localization for mathematical statements.

RDP LoRA uses geometric analysis of LLM representations to identify critical layers for parameter-efficient fine-tuning, achieving superior performance with fewer adapted parameters.

ShadowPEFT introduces a centralized parameter-efficient fine-tuning framework for LLMs that reuses a shared shadow module across layers, offering competitive performance and flexibility.

A data-driven framework using neural operators to discover stability and receptivity properties in physical systems without requiring governing equations.

A humanoid robot learns five distinct gaits using reinforcement learning with a selective adversarial motion prior, achieving better performance and zero-shot sim-to-real transfer.

A deep learning framework for detecting environmental sound deepfakes, outperforming existing methods with a fine-tuned WavLM model.

LASER is a closed-loop framework for active sensing that uses reinforcement learning to guide sensor placement for high-fidelity continuum field reconstruction.

A novel dataset and evaluated models for Romanian legal grammatical error detection and correction to improve legal document accuracy.

TACENR provides task-agnostic contrastive explanations for node representations in graphs, identifying key attribute, proximity, and structural features.

A hybrid anomaly detection framework for medical images using self-supervised learning and Mean Shift Density Enhancement, achieving state-of-the-art results.

A privacy-preserving federated framework for log anomaly detection using parameter-efficient LLMs and differential privacy, matching centralized performance with higher precision.

A framework for evaluating retrieval-augmented generation systems on real-world, redundant corpora by tracking atomic facts and improving LLM data generation.

A lightweight system to detect hallucinations in speech models at inference time using attention maps, improving accuracy and generalizability.

An AI platform that provides real-time feedback on coding consistency for qualitative researchers, reducing interpretation drift.

An LLM agent benchmark for evaluating threat hunting capabilities in cybersecurity, revealing current model limitations.

MATCH is a hybrid control policy for high-precision in-contact manipulation under uncertainty, improving success rates and reducing damage.

FeatGEO optimizes generative answer engine visibility by abstracting webpages into interpretable features, outperforming token-level methods.

A multimodal transformer that predicts properties of experimental metal-organic frameworks by integrating structural identity with X-ray diffraction data for sample-aware analysis.

A framework that reduces decoding redundancy in Diffusion Large Language Models, significantly accelerating inference speed while maintaining generation quality.

AltTrain is a post-training method that modifies the reasoning structure of large reasoning models to improve safety alignment without complex RL, using only supervised finetuning on 1K examples.

An AI agent for social deduction games that integrates memory, social graph analysis, and adaptive communication to win competitions.

A framework and dataset for aligning image editing models with human preferences, improving quality and user satisfaction.

Develop an agent-based system for enhanced multimodal art retrieval and understanding.

A causal inference framework, CD-GNN, that disentangles spurious inductive subgraphs from true causal signals for improved heterophilic graph learning.

This study fine-tunes small reasoning models for Quantum Field Theory using a novel data generation pipeline and publicly releases the data, pipeline, and reasoning traces.

An AI system using wearable data to predict heat stress in construction workers, achieving high accuracy and offering interpretable safety insights.

Enables precise control over LLM behaviors using linear optimal control for safer outputs.

Reinforcement learning improves LLM accuracy and reasoning for disease classification from radiology reports.

Product-of-Experts training reduces dataset artifacts in Natural Language Inference models by downweighting overconfident examples.

Derives achievable QWK ceilings for automated essay scoring from classical test theory, clarifying performance targets and remaining headroom for AES models.

A new benchmark, CulturALL, evaluates LLMs' multilingual and multicultural competence on grounded tasks, revealing significant room for improvement.

A rigorous evaluation of Graph Neural Networks for Bitcoin fraud detection reveals that simpler models outperform GNNs under realistic temporal shifts, with code and a new protocol released for reproducible research.

Adaptive MSD-Splitting enhances C4.5 and Random Forests for skewed continuous attributes, improving accuracy and efficiency.

This paper characterizes AI Engines on FPGAs for extreme-edge scientific computing, providing a metric to determine when they outperform programmable logic.

A study evaluating state-of-the-art methods for detecting data contamination in LLMs, finding current methods unreliable.

This paper adapts LLMs to generate streamliners for Answer Set Programming, achieving significant speedups on benchmark problems.

A human-machine co-boosted active learning framework for efficient and effective bug report identification from GitHub.

A semantic infrastructure protocol for multi-agent LLM systems that enables cross-session cognitive collaboration and memory persistence, running in production.

A new evaluation protocol for continual knowledge graph embedding that addresses overlooked entity interference, leading to more accurate performance assessments.

Tadabur, a large-scale Quran audio dataset with over 1400+ hours from 600+ reciters, offering substantial variation for Quranic speech research.

A human-guided system for recovering AI agents from harmful states, featuring a benchmark and a reward model for preference-aligned remediation.

A hierarchical adversarial fine-tuning framework for vision-language models that improves zero-shot classification robustness across class hierarchies.

OmniMouse scales multi-modal brain models on massive neural data, achieving state-of-the-art performance and revealing data-limited scaling properties.

A novel framework audits LLM similarity scoring, revealing consistent biases and providing a tool to compare model behavior.

An empirical study of agentic security architectures reveals optimal coordination strategies for offensive tasks, balancing coverage and cost.

REVEAL aligns retinal images with clinical risk factors using vision-language models and contrastive learning for early Alzheimer's and dementia prediction, outperforming existing methods.

This paper introduces a 'forecast-necessity' testing framework for interpreting causal discovery in nonlinear time-series models, moving beyond coefficient magnitude for more reliable causal reasoning.

This work proposes an autoregressive sequence modeling approach for healthcare AI that handles missing modalities in patient data, improving diagnostic accuracy and interpretability.

A framework for testing the robustness of tool-integrated AI agents against adversarial manipulation of their tools, revealing critical vulnerabilities.

A novel loss function significantly improves ocean forecasting skill by accounting for variable correlations, offering a better background for data assimilation.

This research introduces a method to enhance Large Language Model robustness against jailbreak attacks by integrating Sparse Autoencoders at inference time, showing significant reductions in attack success rates.

An agentic system leverages the geometric properties of Earth observation embeddings for improved environmental reasoning and query decomposition.

A novel method for error-free training of ML models on MedMNIST datasets, achieving perfect accuracy on most biomedical classification tasks.

A neuro-symbolic pipeline and benchmark for translating natural language reasoning into executable formal logic, enabling more reliable AI reasoning.

ARES is a framework that systematically discovers and mitigates dual vulnerabilities in LLM reward models and core policies, enhancing safety robustness while preserving capabilities.

Develops an agent that automates the process of reproducing and ablating components in virtual cell repositories to identify critical factors.

A replicable methodology for auditing deployed institutional risk models, revealing how disparities in resource allocation emerge and compound across the ML pipeline.

A framework for scaling evaluation-driven discovery loops in language models, demonstrating significant gains across scientific problems.

Analyzes energy impact of centralized vs. decentralized AI for 6G IoT networks, showing distributed models reduce electricity consumption by up to 70% while maintaining predictive accuracy.

A post-hoc calibration method to control LLM distillability via reinforcement fine-tuning, mitigating distillation traps and enabling better knowledge transfer or model IP protection.

Theoretical analysis of adversarial training for Vision Transformers reveals conditions for benign overfitting, improving robustness.

A semi-supervised meta-additive model that automatically identifies informative variables and updates similarity matrices for robust predictions.

A self-improving framework for tabular data generation that uses automated quality signals to iteratively refine language models, outperforming existing methods in fidelity and utility.

Unlock multi-modal reasoning for visual-semantic tasks with LLMs.

Investigating normative conformity in LLMs to understand and potentially control their behavior in multi-agent systems.

A theoretical framework for understanding generalization in neural network training by analyzing optimizer dynamics as fractal attractors.

An execution environment that guarantees AI agent privacy by enforcing user-defined data access permissions without trusting the agent.

Harmful intent is a geometrically recoverable feature in LLM residual streams, detectable with high accuracy across various models and alignment states.

An attention-based multi-modal deep learning model for spatio-temporal crop yield prediction using satellite, soil, and climate data.

A novel tokenization method for symbolic music that represents uniform temporal steps, improving generation quality and efficiency.

Evaluating the trade-offs of small language models under agent paradigms for efficient deployment in resource-constrained settings.

A new benchmark and evaluation framework, Ghost-100, assesses how prompt tone induces hallucination in Vision-Language Models, revealing nuanced sensitivities across different model families and tasks.

A novel method for LLM-based multi-agent systems to infer and track partner traits, improving coordination in complex tasks.

Compares the repeated generation consistency of exercise prescriptions across three LLMs, revealing fundamentally different generative behaviors and implications for reliable deployment.

This paper introduces GOLD-BEV, a framework for dense semantic Bird's-Eye-View mapping of dynamic scenes using synchronized aerial imagery for supervision during training.

Analyzing how answer tokens read reasoning traces in LLMs to improve quantitative reasoning accuracy.

A novel reinforcement learning approach for LLM post-training that adaptively optimizes baseline selection to reduce variance and improve performance across various tasks.

A novel approach to streaming reinforcement learning that aims for predictable per-step changes in function output, leading to state-of-the-art performance.

This research proposes a formal logic (CTLF) with counting worlds semantics to reason about and mitigate bias in series of generative AI outputs, offering theoretical guarantees for fairness.

A framework for evaluating the utility of synthetic trajectory generators and identifying privacy vulnerabilities through adversarial attacks.

A deep learning framework for learning lifted action models from unsupervised visual traces, using mixed-integer linear programming for correction.

A conceptual framework for integrating anomaly detection into agentic AI to proactively manage human activity risks, such as falls, by identifying subtle movement deviations.

A theoretical framework for safety-critical control using online Riemannian optimization with world models, focusing on sample-based density estimation.

A new benchmark dataset and task for evaluating gender-aware morphological generation in multilingual LLMs, revealing significant gaps in current models.

A human-symbiotic agent paradigm for cross-user autonomous cooperation with layered identity, scoped authorization, and action-level accountability.

This paper explores how to design AI in education to foster relational learning, drawing inspiration from Indigenous worldviews and participatory design.

A novel training method for transformers that uses denoising recursion to improve reasoning on complex tasks by learning intermediate refinement paths.

HELM is a model-agnostic framework that enhances long-horizon manipulation by addressing memory, verification, and recovery gaps with an episodic memory module, state verifier, and harness controller.

An IDE plugin that visualizes AI coding history as a graph, enabling developers to explore, compare, and merge AI-generated code states.

Geodesic Tangent Space Aggregation PCA offers a geometry-aware dimensionality reduction method that improves semi-supervised learning.

Quantum-inspired neural networks offer faster and more robust financial forecasting with reduced training times.

This research probes the scientific feasibility assessment capabilities of LLMs, revealing that outcome evidence is more reliable than experiment descriptions for improving accuracy.

A multi-level temporal graph network with local-global fusion is proposed for superior industrial fault diagnosis, outperforming baselines on complex fault scenarios.

A formally verified pipeline using dependent type theory to analyze patents, providing machine-checkable certificates for IP analysis.

Analysis of how LLMs are changing academic peer review, leading to more fluent but less deep evaluations.

This research analyzes the error introduced by convex relaxations in neural network verification, providing theoretical bounds on the divergence between relaxed and original network outputs.

A novel loss function that leverages proxy geographic data to improve the accuracy of Earth observation models with limited labeled data.

A formalization and defense against UI state manipulation attacks on desktop GUI agents, addressing the observation-to-action gap.

A novel intrinsic reward mechanism for world model training that improves exploration by focusing on cumulative prediction error.

A new visualization tool helps users understand the distribution of language model outputs, improving prompt iteration and evaluation.

A novel neural network architecture that learns its own discrete programming language and uses differentiable execution for end-to-end training and test-time adaptation.

A comparative analysis of LLM quantization methods, highlighting reproducibility issues and clarifying differences between RaBitQ and TurboQuant.

Theoretical analysis of Q-learning using a direct stochastic switching system representation for improved convergence bounds.

A geometric framework diagnoses latent diffusion model instability by decoupling local scaling and complexity, identifying 'Geometric Hotspots' as the root cause of unreliable generation.

Identifies and causally links specific neurons in LLMs to citation hallucination, suggesting a method for detection and mitigation.

A first-order logic, called Plausible Logic (PL), designed for defeasible reasoning without probabilities, with 8 reasoning algorithms.

Current LLM-based agents execute scientific workflows but do not exhibit the epistemic patterns that characterize scientific reasoning, leading to unreliable knowledge generation.

This paper critically examines emerging regulations for companion chatbots, proposing a general duty of care to address power asymmetries and control through intimacy.

A conceptual framework for understanding and designing AI co-hosts in livestreaming that adapt bidirectionally with streamers and audiences.