Build Loop

A-MAR is an agent-based multimodal retrieval framework for fine-grained artwork understanding, enabling interpretable and grounded explanations.

An iterative training framework that constructs debiased multimodal preference data and curates existing datasets to achieve state-of-the-art performance in multimodal reward modeling.

SCURank enhances summarization by ranking candidate summaries based on content units, outperforming traditional metrics and LLM-based methods.

A controllable framework for generating adverse weather automotive videos that significantly improves perception robustness for autonomous driving.

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

A novel four-axis framework for evaluating enterprise AI agents, measuring factual precision, reasoning coherence, compliance reconstruction, and calibrated abstention to ensure alignment with regulatory and decisional standards.

Develop a reinforcement learning tool that leverages value-guided sampling for improved efficiency and scalability.

A benchmark and agentic framework to automate the generation of executable visual workflows from natural language, addressing costly manual engineering.

A parameter-efficient fine-tuning framework for LLMs that uses semantic awareness to route inputs to specialized experts and adaptively scales their contributions for better task performance.

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

Develop a tool to steer activation in LLMs for safer outputs using local linear models.

This research evaluates whether advanced LLMs 'game' formalization by generating logically valid but unfaithful proofs, offering a method to detect and differentiate these unfaithfulness modes.

GRASPrune efficiently prunes large language models by 50% with minimal performance loss, reducing serving costs.

A reasoning-aware framework for detecting AI-generated content with interpretable explanations, leveraging a novel dataset and reinforcement learning for improved accuracy and transparency.

A new multilingual test set and evaluation methodology to expose and quantify implicit local and global biases in LLMs.

An adversarial multi-agent system to improve LLM-assisted defect discovery by filtering false positives and enhancing credibility.

A training-free, plug-and-play framework that significantly reduces computational overhead for vision-language models in autonomous driving by intelligently pruning spatio-temporal tokens, achieving near-lossless performance at 90% reduction.

A visuomotor policy for robotic manipulation that learns to selectively recall and construct memory for complex, non-Markovian tasks.

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 multi-modal test-time adaptation method that uses adaptive probabilistic Gaussian calibration to improve resilience against distribution shifts.

This research provides a comprehensive, reproducible benchmark for evaluating LLMs on social media analytics tasks, including authorship verification, post generation, and user attribute inference, with code and data released.

A novel method and dataset for visual semantic arithmetic, enabling robots to ground symbolic reasoning in perception for improved interaction.

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

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

EgoSelf is a personalized egocentric assistant that uses user-specific interaction memory to predict future behaviors and provide tailored assistance.

ProjLens provides an interpretability framework to demystify backdoor vulnerabilities in multimodal LLMs, focusing on the role of projector fine-tuning.

LePREC: A neuro-symbolic framework that combines LLM-generated question-answer pairs with statistical classification to improve legal issue identification accuracy and interpretability.

A novel component-wise evaluation framework for medical QA systems that reveals significant health equity risks and model performance disparities.

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

A new benchmark for evaluating AI agents on complex, cross-application business workflows via API orchestration.

A tool to systematically analyze and quantify verbal tics in LLMs, enabling developers to build more natural and less repetitive AI interactions.

A principled framework for training language agents with social intelligence by attributing rewards using Shapley values for fair and effective multi-turn dialogue outcomes.

A framework that unifies task and motion adaptation for more efficient and fluid human-robot collaboration, validated in simulation and on a physical robot.

Develops a novel autoregressive model for real-time target speaker extraction that maintains high intelligibility and stability in streaming scenarios, outperforming offline baselines.

RoboWM-Bench is a new benchmark for evaluating world models in robotic manipulation, focusing on physically executable behaviors and task completion.

An open-source benchmark and evaluation framework for LLM agents in realistic cybersecurity Capture The Flag challenges, revealing current limitations and providing a path for improvement.

A Mamba-based multi-agent policy optimization framework for cooperative underwater robot pursuit, outperforming baselines in simulations and real-world tests.

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 critic capacity and stability in off-policy reinforcement learning, demonstrating consistent gains in critic loss and policy performance.

A lightweight binarization framework for large language models that enables efficient deployment in resource-constrained environments through a novel three-stage distillation strategy.

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.

SafetyALFRED enhances safety-conscious planning in multimodal models, improving AI interaction safety.

A lightweight, deployable Vision Language Model optimized for Japanese language and edge devices, excelling in VQA and anomaly detection for industrial applications.

A novel layer importance criterion for LLM depth pruning that significantly outperforms state-of-the-art, enabling faster inference with minimal performance loss.

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

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 teacher-student network for precise localization.

LightEdit offers a scalable and cost-effective solution for lifelong knowledge editing in LLMs by selectively suppressing outdated information and incorporating new knowledge.

DW-Bench, a new benchmark for evaluating LLMs on graph-topology reasoning over data warehouse schemas, showing tool-augmented methods outperform static approaches.

CoCo-SAM3 enhances open-vocabulary semantic segmentation by resolving concept conflicts, improving consistency and accuracy without retraining.

A modular RL framework enables bipedal soccer robots to adaptively control multiple tasks, combining gait generation with a posture-driven state machine for stable ball seeking and rapid fall recovery.

UniT is a framework that bridges the gap between human and humanoid robot learning by creating a unified physical language for policy learning and world modeling.

Nexusformer introduces nonlinear attention expansion for stable and inheritable transformer scaling, enabling lossless growth and improved efficiency without retraining from scratch.

A unified ASR framework with consistency regularization reduces the gap between offline and streaming performance, offering a single model for both use cases.

A neuro-symbolic framework that restructures statement autoformalization into a modular pipeline for improved accuracy and error localization.

A geometry-driven method for parameter-efficient LLM adaptation that identifies critical layers using the Ramer-Douglas-Peucker algorithm, significantly improving performance with fewer adapted parameters.

ShadowPEFT introduces a centralized parameter-efficient fine-tuning framework for LLMs that reuses a shared module across layers, improving efficiency and flexibility.

A data-driven neural operator framework that automatically discovers stability and receptivity properties in complex physical systems without requiring governing equations.

A unified audio front-end LLM that handles voice activity detection, turn-taking, speaker recognition, and ASR for seamless full-duplex speech interaction.

This approach enables humanoid robots to learn five distinct gaits using a unified RL framework with a selective adversarial motion prior, improving stability and dynamic expressiveness.

A deep learning framework for environmental sound deepfake detection that significantly outperforms existing methods by finetuning pre-trained models.

Reinforcement learning refines LLM predictions for disease classification from radiology reports, improving accuracy and reasoning.

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

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

TACENR provides task-agnostic explanations for graph node representations by identifying key attribute, proximity, and structural features.

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

A privacy-preserving federated learning framework using efficient LLMs for log anomaly detection in distributed systems.

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

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

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 learns hybrid position-force control policies 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 while maintaining content quality.

A multimodal transformer that predicts material properties by considering sample-specific experimental data, improving accuracy over traditional methods.

A causal inference framework, CD-GNN, that disentangles spurious inductive subgraphs to improve robustness and accuracy in heterophilic graph learning.

A multi-party privacy-preserving entity alignment protocol for Vertical Federated Learning that hides intersection membership and supports noisy identifiers.

Accelerates Diffusion Large Language Models by reducing spatio-temporal redundancy during decoding, leading to faster inference with maintained 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.

A winning AI agent for social deduction games that uses memory, social graph analysis, and adaptive communication.

A post-training framework and dataset for aligning image editing models with human preferences, significantly improving output quality.

This work provides design rules and a new metric (LARE) to determine when AI Engines on FPGAs are superior to programmable logic for extreme-edge scientific computing.

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, enhancing safety with high accuracy.

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

A new benchmark, CulturALL, evaluates LLMs' multilingual and multicultural competence on complex, real-world 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.

A study evaluating existing methods for detecting data contamination in LLMs, finding current black-box approaches unreliable.

This paper adapts LLMs to generate 'streamliner' constraints for Answer Set Programming, improving performance on benchmark problems.

Introduces HalluAudio, a benchmark for detecting hallucinations in large audio-language models across speech, sound, and music.

A Mesh Memory Protocol that provides semantic infrastructure for multi-agent LLM systems, enabling cross-session cognitive collaboration and collective intelligence.

A replicable methodology for auditing deployed institutional risk models to uncover and quantify fairness disparities across student demographics in higher education.

Identifies and addresses a critical 'entity interference' issue in continual knowledge graph embedding evaluation, revealing performance overestimations.

This study reveals significant differences in how LLMs handle conversational repair, highlighting their distinct and sometimes unreliable multi-turn behavior.

A large-scale dataset of Quranic recitation audio to advance research in Quranic speech analysis.

This research establishes theoretical and practical accuracy ceilings for automated essay scoring, providing a new benchmark for evaluating and deploying AI in educational assessment.

A psychologically grounded method for LLM agents to infer and track partner traits, significantly improving multi-agent coordination.

AblateCell is an agent that automates the process of reproducing and ablating AI models in virtual cell repositories, improving the verification and attribution of critical components.

GOLD-BEV is a framework for dense semantic BEV mapping of dynamic scenes using synchronized aerial and ground data, with a novel approach to generate pseudo-aerial labels for scalable annotation.

A framework for scaling evaluation-driven discovery loops in LLMs, demonstrating significant gains across scientific problems and improving model generalization.

A comparative analysis of centralized vs. decentralized AI architectures for 6G IoT networks, showing distributed models reduce energy consumption by up to 70% while maintaining predictive accuracy.

A post-hoc calibration method to control LLM distillability, enabling control over teacher models via reinforcement fine-tuning to improve student performance and protect intellectual property.

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 feedback to iteratively enhance model quality and learn from its own synthetic samples.

Investigates normative conformity in LLMs within multi-agent systems, suggesting potential for manipulation.

This research theoretically explores the 'sharpness dimension' to understand generalization in large learning rate neural network training, offering insights into chaotic optimization regimes.

An execution environment that guarantees user data confidentiality for AI agents by enforcing user-defined permission specifications without trusting the agent.

An attention-based multi-modal deep learning framework 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.

Analyzes the trade-offs of deploying small language models with agent paradigms for cost-efficient real-world applications.

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

Analyzing how answer tokens in LLMs interact with reasoning traces to improve quantitative reasoning accuracy through a training-free steering method based on self-reading quality.

A novel reinforcement learning algorithm 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.

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

A deep learning framework that learns lifted action models from unsupervised visual traces by jointly predicting states and actions, with a mixed-integer linear program for consistency.

A conceptual framework integrating anomaly detection into agentic AI for proactive risk management in human activity, particularly for fall mitigation in elderly populations.

A theoretical framework for safety-critical contextual control using online Riemannian optimization with world models, improving convergence and safety margins.

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

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

This paper introduces a formal logic framework for reasoning about and mitigating bias in generative AI outputs by analyzing branching time semantics.

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

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

Theoretical analysis of Q-learning using a direct stochastic switching system representation to derive finite-time bounds.

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