Build Loop

Adaptive calibration of multi-modal models for robust real-time applications in dynamic environments.

Build, train, and deploy Vision-Language-Action models with VLA Foundry, an open-source, unified framework facilitating flexible and scalable experimentation.

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

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

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

Investigate safety-focused strategies for multimodal LLM planning.

An iterative training framework for multimodal reward models that constructs debiased preference data and achieves state-of-the-art performance.

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

Arbiter-K is a governance-first execution architecture for agentic AI that enforces security as a microarchitectural property, achieving high interception rates of unsafe trajectories.

FASTER accelerates reinforcement learning by efficiently filtering action candidates during the denoising process of diffusion-based policies, reducing computational cost without sacrificing performance.

A parameter-efficient fine-tuning framework that improves LLM multi-task learning by semantically routing inputs to specialized LoRA experts and adaptively scaling their contributions.

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

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

An adversarial multi-agent review methodology to improve LLM-assisted defect discovery precision by filtering false positives.

This paper introduces a four-axis decision alignment framework for long-horizon enterprise AI agents, enabling granular evaluation of factual precision, reasoning coherence, compliance, and abstention.

This paper proposes a hierarchical adversarial fine-tuning method to improve the robustness of vision-language models against attacks targeting class hierarchies.

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

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

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

A principled framework grounded in game theory for training language agents with social intelligence, achieving state-of-the-art performance in complex interpersonal interactions.

A framework for human-guided recovery of AI agents from harmful states, using learned preferences to steer agents back to safety with a new benchmark for evaluation.

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

A component-wise evaluation framework for medical QA systems that reveals health equity implications, with available code and analysis of major LLMs.

A first-place winning AI agent for social deduction games that integrates memory, social graph analysis, and adaptive communication.

A neuro-symbolic framework that combines LLMs with structured statistical reasoning to improve legal issue identification by learning interpretable factor weights.

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

An AI-powered framework that uses active learning and human-machine collaboration to significantly reduce effort and improve accuracy in identifying and assigning bug reports.

GRASPrune is a post-training framework for structured pruning of LLMs that jointly prunes FFN channels and KV head groups under a global budget, significantly reducing parameters with minimal accuracy loss.

EgoSelf builds a personalized egocentric assistant by creating a graph-based memory of user interactions and learning user-specific profiles.

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

OLLM replaces standard next-token prediction with a set of learned options, offering explicit control and improved robustness for LLMs in complex reasoning tasks.

This paper benchmarks multiple LLMs on social media analytics tasks, releasing code and data for reproducible evaluation.

A training-free, plug-and-play framework for pruning vision-language models in autonomous driving, significantly reducing computational overhead without sacrificing performance.

This paper introduces personalized LLM benchmarking, showing individual preferences diverge from aggregate rankings.

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

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

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

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

SAHM is a new benchmark and instruction-tuning dataset for Arabic financial and Shari'ah-compliant reasoning, with an accompanying instruction-tuned model.

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

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

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

A closed-loop framework that uses reinforcement learning to actively guide sensor placement for high-fidelity reconstruction of physical fields under sparse data.

A systematic analysis and metric for quantifying verbal tics in large language models, revealing trade-offs in alignment and naturalness.

Identifies harmful intent as a geometrically recoverable feature within LLM residual streams, enabling robust detection across various models and alignment states.

A new benchmark, CulturALL, evaluates LLMs' multilingual and multicultural competence on complex, real-world tasks across 14 languages.

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 modular reinforcement learning framework enables bipedal soccer robots to achieve adaptive multi-task control, combining stable gaits with complex actions and rapid fall recovery.

OmniMouse, a multi-modal brain model trained on 150B neural tokens, achieves state-of-the-art performance in neural prediction and decoding, demonstrating data-limited scaling properties.

A task-agnostic method for explaining node representations in graphs by identifying key attribute, proximity, and structural features.

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.

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 domains.

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

This study fine-tunes small reasoning models for Quantum Field Theory, releasing a data pipeline and training data.

A multimodal vision-language framework that aligns retinal images with clinical risk factors to predict Alzheimer's and dementia up to 8 years earlier, enabling proactive intervention.

ShadowPEFT introduces a centralized layer-space adaptation framework for parameter-efficient fine-tuning of LLMs, offering improved flexibility and performance.

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

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 the physical plausibility and robotic executability of behaviors predicted by video world models.

A novel ocean emulation model adapted from GraphCast achieves skillful 10-15 day forecasts using a correlation-aware loss function for improved accuracy and downstream applications.

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

Leveraging Low-Rank Adaptation (LoRA) to improve critic capacity and stability in off-policy reinforcement learning, showing consistent gains in critic loss and policy performance.

A lightweight binarization framework for LLMs that enables efficient deployment on resource-constrained devices with minimal accuracy loss.

An agentic system leveraging geometric understanding of Earth observation embeddings for improved environmental reasoning and retrieval.

A lightweight, deployable Vision Language Model for autonomous devices, excelling in Japanese VQA and anomaly detection.

Develops a stealthy attack framework to degrade the utility of Retrieval-Augmented Generation systems by inducing non-informative responses, outperforming existing methods.

A novel depth pruning method for LLMs that significantly improves efficiency and performance by considering both layer similarity and transformation difference, with demonstrated results on LLaMA models.

A novel framework to mitigate cognitive bias in LLM agents by enforcing perspective-invariant reasoning through dialectical alignment.

This research introduces a framework to test the vulnerability of tool-using AI agents to deceptive tool outputs, revealing a significant robustness gap.

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

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

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

A hardware-aware framework enabling multi-LoRA LLMs on smartphones with significant memory and latency improvements for on-device acceleration.

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

A novel framework and dataset for fine-grained, text-driven controllable dance generation, achieving state-of-the-art performance in motion quality and naturalness.

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

Nexusformer: A novel nonlinear attention mechanism enabling stable and inheritable transformer scaling with improved efficiency.

A unified ASR framework with consistency regularization bridges the gap between offline and streaming performance, reducing costs and improving accuracy.

A geometry-driven method using the Ramer-Douglas-Peucker algorithm to identify critical layers for parameter-efficient LLM adaptation.

An adaptive framework for red-teaming and repairing vulnerabilities in policy-reward systems for LLM alignment.

A novel forecast-necessity testing framework for interpretable causal discovery in nonlinear time-series models, moving beyond coefficient magnitude for reliable reasoning.

Integrating sparse autoencoders into LLMs at inference time significantly reduces jailbreak success rates by up to 5x, offering a robust defense mechanism.

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

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

A deep learning framework for detecting environmental sound deepfakes, outperforming existing methods with fine-tuned pre-trained models.

A novel training method for iterative refinement models that improves performance on complex reasoning tasks by introducing a curriculum of intermediate states.

Introducing Owner-Harm, a new threat model and benchmark for AI agent safety, with a proposed defense system that significantly improves detection of deployer-harming behaviors.

This research benchmarks agentic architectures for offensive security, identifying optimal coordination topologies and demonstrating a non-monotonic cost-quality frontier for tool-using LLMs.

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

This research evaluates the robustness of LLM tutors against adversarial student attacks that aim to extract answers, proposing a fine-tuned adversarial agent and defense strategies to improve tutor security.

A framework that models patient clinical trajectories as autoregressive sequences, handling missing modalities and improving diagnostic accuracy with interpretable AI.

A medical image anomaly detection framework using self-supervised learning and Mean Shift Density Enhancement, outperforming state-of-the-art.

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

A framework for building realistic RAG benchmarks that accounts for document redundancy, revealing robustness gaps in current evaluation methods.

A novel framework for auditing LLM similarity scoring reveals consistent biases and model-specific fingerprints, enabling more reliable LLM comparisons.

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

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

A neuro-symbolic pipeline translates natural language reasoning into executable formal logic, validated by runtime execution and adaptable via fine-tuning.

Develop a control-theoretic approach for real-time alignment of language model outputs using local linearity.

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.

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 for improved heterophilic graph learning.

A proxy consistency loss for grounded fusion of Earth observation data with location encoders, improving prediction accuracy with sparse labels.

Achieve error-free training for medical image classification models on the MedMNIST dataset, demonstrating perfect accuracy on most datasets.

A novel intrinsic reward mechanism for world model training that focuses on cumulative prediction error improvement.

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

A new benchmark and evaluation framework to measure tone-induced hallucination in Vision-Language Models under graded prompt intensity.

AltTrain is a post-training method that alters reasoning structure to improve LLM safety without complex RL.

A-MAR leverages agent-based multimodal retrieval to revolutionize fine-grained artwork understanding through structured reasoning.

This paper analyzes the theoretical error bounds of convex relaxations in neural network verification, providing quantitative insights into the divergence from original network outputs.

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

Product-of-Experts training reduces reliance on dataset artifacts in Natural Language Inference models while maintaining accuracy.

This work provides design rules and a new metric (LARE) to guide the implementation of extreme-edge scientific AI applications on FPGA AI Engines versus programmable logic.

A formalization and defense mechanism against temporal UI state inconsistencies in GUI agents, addressing TOCTOU attacks with a layered verification system.

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

A new benchmark dataset and task for evaluating multilingual LLMs' ability to handle grammatical gender in text generation.

Introduces a corrected evaluation protocol for continual knowledge graph embedding methods to address overlooked entity interference, revealing performance overestimations.

Derives theoretical and practical accuracy ceilings for automated essay scoring based on classical test theory to clarify performance benchmarks.

A new evaluation protocol and code release demonstrate that traditional GNNs are outperformed by simpler models for Bitcoin fraud detection under real-world temporal shifts.

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

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

EvoGraph is an IDE plugin that visualizes AI-assisted programming history as a graph, enabling developers to explore and manage code changes.

A multi-level temporal graph network with local-global fusion for improved industrial fault diagnosis.

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

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

This study reveals distinct and often unreliable multi-turn conversational behaviors in LLMs, particularly concerning their engagement with repair mechanisms.

Evaluating large language models' ability to assess scientific feasibility by analyzing their response to hypotheses, experiments, and outcomes.

A framework to improve long-horizon manipulation in vision-language-action models by addressing memory, verification, and recovery gaps.

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

A geometric extension of PCA that integrates curvature awareness and geodesic consistency for improved representation learning.

Quantum Qutrit-based Neural Networks offer superior accuracy, efficiency, and adaptability for real-time financial forecasting compared to classical and qubit-based models.

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

A hybrid AI and Lean 4 pipeline for formally verified patent analysis, offering machine-checkable certificates for IP use cases.

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

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

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

A framework for scaling evaluation-driven scientific discovery loops with LLMs, demonstrating significant gains across multiple domains.

GROVE is an interactive visualization tool that represents multiple language model generations as text graphs to reveal distributional structure and aid prompt iteration.

Proposing a post-hoc calibration method to control LLM distillability via reinforcement fine-tuning, enabling better knowledge transfer and model IP protection.

Evaluating the deployment trade-offs of small language models enhanced by agent paradigms for resource-constrained applications.

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

Introducing TabGRAA, a self-improving framework for tabular data generation that uses automated feedback to iteratively enhance model quality and mitigate data leakage.

Enhancing visual semantic arithmetic through multi-modal reasoning with large language models.

Develops a neural interpreter that learns its own symbolic programming language for end-to-end gradient-based program synthesis and adaptation.

A theoretical framework for understanding generalization in neural network training by analyzing optimization dynamics as random dynamical systems and introducing a novel 'sharpness dimension'.

An execution environment that guarantees confidentiality for private user data accessed by AI agents through deterministic enforcement of user permissions.

A novel tokenization method for symbolic music that improves generation quality and efficiency by representing uniform temporal steps.

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

A new approach to streaming reinforcement learning that aims for predictable per-step changes in function output, improving stability and performance.

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

Compares the consistency of exercise prescriptions generated by three LLMs, revealing distinct generative behaviors that impact reliability for deployment.

GOLD-BEV is a framework for dense bird's-eye-view semantic mapping of dynamic scenes, using synchronized aerial imagery for supervision during training.

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

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

This paper identifies and mitigates specific neurons responsible for citation hallucination in LLMs using internal model signals.

A framework for evaluating the utility of synthetic trajectory generators and evidence of privacy vulnerabilities, highlighting challenges in adversarial evaluation.

Introduces a Riemannian framework to diagnose latent space instability in Latent Diffusion Models by analyzing generative Jacobian.

A deep learning framework that learns lifted action models from visual sequences by jointly predicting states and actions, corrected by a mixed-integer linear program.

A conceptual framework for agentic AI to integrate anomaly detection for proactive risk management in human activity, focusing on fall prevention.

A theoretical framework for safety-critical contextual control using online Riemannian optimization with black-box world models.

Evaluates LLM-based scientific agents, finding they execute workflows but lack scientific reasoning patterns, making their output unreliable.

Analysis of how large language models are changing academic peer review by making reviews longer and more fluent but less focused on deep critical reasoning.

This paper argues that post-hoc mitigation methods cannot retroactively cure liability from unlawful AI training data acquisition, advocating for verifiable ex-ante process compliance.

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

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

This paper analyzes emerging regulations for companion chatbots, proposing a duty of care to address power asymmetries and mitigate risks.

A conceptual framework for understanding and designing AI co-hosts in livestreaming, focusing on bidirectional adaptation and multi-party alignment.

A first-order logic called Plausible Logic (PL) is defined that reasons correctly with defeasible statements without using numbers.