Build Loop

Developing GUI agents that learn through reinforcement learning to become digital inhabitants, addressing long-horizon credit assignment and safety.

KellyBench is a new benchmark and API for evaluating long-horizon sequential decision-making in dynamic environments like sports betting markets.

Enabling instruction-guided poetry generation in Arabic and its dialects with a large-scale dataset and fine-tuned LLMs, assisting users in creative writing.

NeocorRAG enhances Retrieval-Augmented Generation by optimizing retrieval quality through evidence chains, achieving state-of-the-art performance with significantly fewer tokens.

ValuePlanner introduces a hierarchical cognitive architecture for embodied agents, enabling self-directed behavior by decoupling high-level value reasoning from low-level action planning.

A high-quality corpus and hybrid framework for summarizing and interpreting application privacy policies, outperforming large language models in readability and reliability.

TeCoD boosts Text-to-SQL accuracy by converting query patterns into reusable templates and enforcing them during SQL generation, reducing latency and improving execution accuracy.

Automated diagnostics for transformer models that detect, categorize, and pinpoint root causes of faults, improving repair accuracy for practitioners.

This research develops a deep learning system to automatically segment peritoneal cancer index regions from CT scans, enabling non-invasive assessment and approaching interobserver agreement.

A lightweight, real-time crack classification model for UAV bridge inspections that balances accuracy, speed, and robustness.

D3-Gym provides verifiable environments and a dataset for training AI agents on real-world scientific discovery tasks, significantly improving model performance.

MCPHunt is a benchmark framework for evaluating cross-boundary data propagation in multi-server agents, identifying vulnerabilities and enabling prompt-level mitigation strategies.

A language-agnostic benchmark generator for RTL code completion tasks, enabling controlled and scalable evaluation of LLM capabilities in hardware design.

A lightweight module accelerates LLM-based recommendation inference by improving speculative decoding with position-aware signals.

A new benchmark for evaluating autonomous GUI agents on complex, cross-application professional workflows, revealing significant performance gaps in current leading models.

A dynamic agentic framework for time series forecasting that refines predictions through iterative planning, action, and reflection, leveraging specialized LLMs and ensemble methods.

A comprehensive SHACL validation framework for TMF Intent Ontologies to ensure correctness of network intents before deployment.

A hierarchical alignment framework for generating accurate radiology reports from medical images by matching visual features to report structure.

SpecVQA is a benchmark for spectral understanding and visual question answering in scientific images, aiming to improve multimodal models' capabilities in scientific data analysis.

MetaSymbO is a multi-agent framework for language-guided metamaterial discovery, using symbolic latent evolution to generate valid and novel microstructures.

TransVLM is a vision-language framework for detecting continuous shot transitions in videos by incorporating optical flow, achieving state-of-the-art performance and deployed to production.

This research introduces an inference-time feedback mechanism for tool-calling agents, enabling proactive error correction and improving agent performance without retraining.

COHERENCE: A benchmark for evaluating fine-grained image-text alignment in interleaved multimodal contexts, crucial for real-world document understanding.

AutoSurfer is a web agent training data generator that uses breadth-first exploration and guided task synthesis to comprehensively cover websites and improve LLM performance on complex web tasks.

InteractWeb-Bench, a new benchmark and interactive environment for evaluating multimodal agents in website generation under realistic, ambiguous user instructions.

A novel attack that steals sensitive secrets from local LLM fine-tuning by exploiting supply-chain model code backdoors, bypassing existing defenses.

An LLM-based agent system that autonomously performs end-to-end scientific discovery on a real optical platform, identifying a novel physical mechanism.

Intent2Tx benchmarks LLMs for translating natural language intents into Ethereum transactions, revealing critical gaps in reasoning-to-execution for Web3 agents.

A framework for confident LLM migration in production systems, using Bayesian calibration of automated metrics against human judgment for enterprise AI services.

A live benchmark for LLM agents that evaluates their ability to complete evolving real-world workflows with verifiable execution traces.

PhyCo introduces controllable physical priors into video generation using a physics-supervised diffusion model and VLM-guided reward optimization.

A framework for graph few-shot learning that uses hyperbolic space and denoising diffusion to improve representation learning and generalization.

A survey and practical guide for building and evaluating LLM systems across the entire academic peer review workflow.

An end-to-end governance framework for EHR-embedded AI agents, demonstrating significant performance improvements and reduced error rates through continuous monitoring and controlled experimentation.

TEA Nets is an open-source Python library combining AI and cognitive network science to model targets, events, and actors in text, enabling interpretable emotion detection and semantic analysis.

A multimodal retrieval-augmented generation system for medical question answering that reasons over document images, achieving state-of-the-art accuracy.

MIFair is a unified framework for bias assessment and mitigation based on mutual information, addressing intersectionality and multiclass fairness challenges.

Leveraging LLMs to refine graph structures for EEG seizure diagnosis, improving accuracy and interpretability by removing noisy connections.

A self-supervised, feature-centric contrastive framework for learning informative representations from challenging tabular remote sensing data.

DALS is a unified optimizer framework that integrates phase-adaptive scheduling and depth-aware scaling for efficient and effective model training across diverse tasks.

Eywa is a framework enabling language models to orchestrate and reason over diverse scientific foundation models for complex, multi-modal tasks.

BrainDINO, a self-supervised foundation model trained on millions of brain MRI slices, generalizes across diverse clinical tasks with minimal labeled data.

An open-source, low-cost, and scalable piezoresistive tactile sensing solution for robotic end-effectors that enables advanced tactile learning pipelines.

This research proposes a fundamental shift in agent memory systems, moving beyond simple lookup to true generalization by learning abstract rules, addressing limitations in current AI capabilities and security.

An open-source toolkit for automated vibrational analysis and structural remediation of crystalline materials using machine-learned potentials, enabling faster discovery of stable polymorphs and functional materials.

A schema-grounded AI memory system that reliably extracts and validates factual information for agents, moving beyond simple text retrieval.

A multi-agent framework with retrieval augmentation and structured reasoning for robust multimodal stance detection.

WaferSAGE is an LLM-powered framework for wafer defect analysis using synthetic data and reinforcement learning, enabling on-premise deployment of small, specialized models.

A post-optimization method to significantly reduce LoRA parameters by 75-90% while preserving performance, enabling more efficient fine-tuning of large models.

A novel method for targeted knowledge discovery and fuzzy logic update in deep neural networks to improve robust image recognition without explicit concept labels.

High-signal data filtering for sample-efficient German language modeling, achieving state-of-the-art results with significantly fewer tokens.

Language models and numerical optimizers collaborate to systematically improve mechanical linkage designs by exploring topologies and fitting parameters.

A hybrid control architecture for autonomous robots that integrates LiDAR-based latent dynamics with MPPI for dense, dynamic crowd navigation.

A method to debias reward models for LLMs at inference time by intervening on specific neurons, improving alignment without performance trade-offs.

BoostLoRA enhances parameter-efficient fine-tuning by iteratively merging low-rank adapters to grow effective rank, achieving state-of-the-art performance without inference overhead.

A framework for video moment retrieval that improves accuracy by learning relationships between multiple relevant video segments.

HGUL: A framework for robust representation learning on heterogeneous graphs with heterophily by jointly handling noisy structures and learning adaptive affinities.

A new metric for LLM evaluation that detects reliable changes in model performance beyond aggregate scores, revealing bidirectional improvements and deteriorations.

A methodology for systematically engineering LLM agents with subject matter experts, developers, and helper agents, improving development efficiency and complex-query performance.

Develops an activation-level detection system for multi-turn LLM prompt injection attacks, achieving high accuracy across multiple model families.

Machine collective intelligence autonomously discovers explainable governing equations from data, outperforming deep learning in extrapolation and parameter efficiency.

This research introduces the 'Inverse-Wisdom Law' and 'Architectural Tribalism' to explain why agentic swarms can prioritize internal agreement over factual accuracy, proposing a 'Heterogeneity Mandate' for resilient architectures.

A schema-agnostic evaluation framework for production Text-to-SQL systems that provides continuous monitoring and improvement feedback.

An end-to-end LLM framework automating threat detection, query generation, and incident resolution in Security Operations Centers, reducing triage time from hours to minutes.

PRISM uses on-policy distillation to improve multimodal reinforcement learning by aligning policy with supervision distribution, reducing drift from supervised fine-tuning.

A generative multimodal model of human physiology that forecasts individual health trajectories and simulates interventions for personalized medicine.

An empirical study and benchmark dataset analyzing how generative AI disrupts web search, comparing Google Search, Gemini, and AI Overviews to understand information presentation and source differences.

Reframing clinician overrides as implicit preference signals for clinical AI, enabling robust learning in value-based care settings.

A multi-layer adaptive aggregation method for Byzantine-robust federated learning that defends against complex attacks without requiring server-side data.

A framework translating natural language to Answer Set Programming enables LLMs to perform task-agnostic nonmonotonic reasoning through automated self-correction.

Evaluating LLM agents for scientific visualization tasks, comparing interaction paradigms and modalities for optimal workflow generation.

Introducing a unified framework where multi-agent systems performing local free-energy minimization implicitly implement a stochastic game.

A hybrid engine that uses LLM agents and a Protocol-Aware DSL with templates to generate UVM testbenches and sequences for IC verification.

TopBench: A new benchmark for evaluating LLMs on implicit prediction and reasoning over tabular data.

A controlled environment and metric space to rigorously assess and improve the out-of-distribution generalization of neural program synthesis models.

Pragmos: A prototype system for collaborative, explainable process modeling using LLMs and specialized tools.

This paper presents a configurable, grid-aware Agent-Based Model in Python for analyzing electric vehicle charging systems, integrating heterogeneous EV behavior and facility-level power dynamics.

A dataset mapping LLM mathematical reasoning and confidence across simulated student personas, integrating self-efficacy and anxiety for safer AI tutors.

Identifies distinct profiles of AI dependency among Filipino students, highlighting risks to academic competencies and advocating for AI literacy.

This research highlights the critical impact of prompt optimization on LLM evaluation, suggesting a need for more dynamic and model-specific assessment methods.

An adaptive inference control framework for black-box LLM services that balances response quality with computational cost using verifiable observations.

An all-MLP framework for multivariate time series forecasting that uses iterative refinement and external attention, optimized by Harris Hawks Optimization for adaptive dropout tuning.

Analyzes factors leading to non-development or abandonment of AI systems, identifying levers beyond ethical concerns.

A new method for reliable circuit diagram to RTL code generation that addresses the 'Mirage' phenomenon and achieves strong visual grounding.

A framework for auditing financial NLP benchmarks to ensure reliable model selection and deployment by addressing measurement risk.

A formal framework for runtime delegation safety in hierarchical multi-agent systems, balancing safety and efficiency.

Exploring the co-evolution of agents and institutions to model societal dynamics.

Evaluating the emotional nuance preservation of small language models in machine translation using a fine-grained emotion dataset.

Mechanized proofs for structural governance in cognitive workflow systems, including coinductive safety predicates and governance invariance theorems.

This research reveals that LLM political bias audits are significantly influenced by sycophancy, demonstrating that bias is a response profile rather than a fixed ideology.

Develops cognitive models to understand and improve human comprehension of AI explanations, with available code for research.

A type-aware methodology for heterogeneous attribute completion in graphs, improving downstream learning by addressing information asymmetry.

RoundPipe is an open-source Python library that enables efficient fine-tuning of large language models on multiple consumer GPUs by breaking weight binding constraints.

A novel LLM architecture that separates reasoning and non-reasoning modes using dedicated experts to eliminate leakage and improve conciseness in non-thinking responses.

A formal framework for analyzing the structural gap in AI system governance, proposing coterminous governance as a testable criterion.

A synthetic corpus and interactive platform for analyzing LLM discourse across diverse human personas and societal topics to audit bias and social sensitivity.

A methodological evolution graph for AI research that automatically identifies and links research methods to enable AI-driven scientific discovery.

Guidelines for creating adversarial, difficult, and legible benchmark tasks for terminal-agent evaluations to improve LLM capability assessment.

This framework explains uncertainty in conformal predictions by localizing calibration, enhancing interpretability without altering the predictor.

Lightweight distillation of foundation models for edge-deployable livestock monitoring, enabling longitudinal visual analytics with significantly reduced computational footprint.

An AI-powered automaton that orchestrates research software development by coupling ideation, implementation, and documentation to prevent hallucination and desynchronization.

An event-driven agent framework that optimizes compute by using small policies by default and escalating to larger models only when risk is detected.

RecGen: a generative framework for reconstructing complex 3D multi-object scenes from sparse observations with state-of-the-art performance.

This paper presents a physically constrained equivariant model that predicts density matrices for accelerated self-consistent field workflows in computational chemistry, reducing iteration steps by up to 81%.

A decentralized framework for trustworthy AI services, enabling robust, scalable, and private verification of large reasoning models through hierarchical auditing and consensus.

This research offers a method to improve LLM instruction following by decoupling reasoning patterns from specific tasks, potentially leading to more controllable and faithful AI.

A foundation model for predicting Alzheimer's disease conversion from Mild Cognitive Impairment, outperforming traditional methods in data-limited settings.

An AI system that automatically generates end-to-end machine learning pipelines from data and natural language goals, improving efficiency and robustness.

A gated hybrid collaborative filtering framework that integrates review semantics for improved ranking in recommender systems.

A domain-specific language and agent for generating spatially accurate 3D indoor scenes from natural language, improving fidelity and plausibility.

A risk-sensitive bandit controller for LLM coding agents that learns when to use external memory to avoid unsafe injections.

A scalable methodology for creating synthetic computer environments and simulating long-horizon productivity tasks to train agents for improved performance.

A Vision-Language-Action foundation model for robotics that reformulates pretraining using goal-conditioned reinforcement learning to improve task progress understanding.

An end-to-end system that translates economic intuitions into executable computational experiments, outperforming generic LLMs in research idea generation.

A self-evolving framework that autonomously discovers, refines, and selects context-specific skills for language models without human supervision.

A survey and taxonomy of graph-based world models for AI agents, exploring relational inductive biases for improved reasoning and planning.

An anchored-curriculum framework where a unified policy learns to generate verifiable problems, solve them, and self-improve using feedback.

A novel rule-generation perspective for estimating LLM compositionality, addressing explainability and data leakage issues.

Develop a native file format for enhanced document and knowledge traversal in digital ecosystems.

Introduces a lightweight agent architecture with explicit state dynamics to model clinical concern trajectories and provide pre-escalation signals.

A leakage-safe survival framework for predicting Alzheimer's disease progression and dynamic tracking using temporal Transformers.

A new approach to procedural tasks using in-context prompting for LLM self-orchestration, outperforming traditional agent orchestration frameworks.

A transparent runtime that bridges the agent-OS semantic gap for efficient and correct checkpoint/restore of agent sandboxes.

A training-free framework for tunnel defect inspection and engineering interpretation using visual recalibration and entity reconstruction.

An agentic framework that constructs knowledge graphs from AI policy documents to reason about policy compliance.

Agentic AI optimizes traffic signals in real-time using a digital twin and existing traffic management APIs.

Consumer attitudes towards AI in digital health are mixed, with preference for AI summaries based on communication quality and human governance.

A neuro-symbolic framework for scalable, explainable, and verifiable rule synthesis in safety-critical domains using LLMs.

A novel framework unifies diffusion-based generation and random structure search for efficient and diverse discovery of molecular and crystal structures.

Enables on-demand persona-based agent generation to dynamically tailor multi-agent workflows to user needs and task contexts.

A priority-aware federated reinforcement learning system for autonomous vehicles to optimize lane changes based on destination urgency.

This paper introduces a method to measure and improve role fidelity in LLM-based political statement analysis systems, revealing failure modes and outperforming existing models.

A new diffusion model approach for generating continuous-time, continuous-space stochastic processes like videos and weather forecasts, conditioned on partial observations.

Auditing frontier vision-language models for trustworthiness in medical VQA, identifying grounding failures and domain adaptation needs.

Investigating how visual priming affects the cooperative behavior of Vision-Language Models in the Iterated Prisoner's Dilemma.

Demonstrates that in-context examples can suppress scientific knowledge recall in LLMs, shifting computation towards empirical pattern fitting rather than knowledge-driven derivation.

Developing local LLM agents to enhance their capability for Linux privilege escalation attacks.

Examines factors influencing pre-service teachers' intention to use AI-enabled educational tools, finding personal motivation more impactful than institutional factors.

A protocol for verifying and deploying LLM-generated reward hypotheses in reinforcement learning based on policy competence and training phase.

Leveraging LLMs for Design Structure Matrix modularization without domain knowledge, achieving near-reference quality in engineering design.

Identifies vulnerabilities in cross-modal encoders by detecting single hub texts that disproportionately achieve high similarity scores with many images, impacting retrieval and evaluation.

Investigating the inconsistent alignment persona of large language models after fine-tuning on misaligned data.

A pattern language for resilient visual agents, separating fast reflexes from slow supervision in enterprise ecosystems.

This paper introduces an energy-geography framework for geo-distributed AI inference, modeling computation placement as a constrained optimization problem considering electricity prices, carbon intensity, and latency.

A framework for governing LLM updates in the software supply chain, proposing production contracts, risk-category-based testing, and compatibility gates to manage behavioral drift.

Flow Map Reward Guidance offers a training-free, single-trajectory method for generative model guidance, achieving state-of-the-art results with significantly fewer steps.

Unsupervised machine learning for anomaly detection in soil heavy metal contamination to enable targeted environmental risk assessment.

A benchmark for evaluating if LLMs can safely recover utility by clarifying user intent in multi-turn conversations.

Investigating the use of LLMs and Diffusion Models for procedural content generation of trading cards to create personalized and dynamic designs.

This paper proposes a protocol and empirical analysis for evaluating the 'epistemic guardrails' of AI reading assistants to understand how they manage interpretive work.

Develops a theoretical framework to understand how sparse autoencoders capture concept manifolds, identifying suboptimal recovery and motivating new interpretability methods.

This paper explores the conceptual foundations and future directions of agentic reinforcement learning within large language models, focusing on autonomous agents capable of complex reasoning and planning.

Analyzing how uncertainty in data affects the execution and outcomes of multi-agent workflows.

A layered review of security risks and defense strategies for autonomous agent frameworks, using OpenClaw as a case study.

Sampler-Robust Optimization (SRO) optimizes decisions against the worst-case sampler induced by perturbing learned generative models.

This work introduces self-evolving software agents that combine BDI reasoning with LLMs to autonomously update their goals, reasoning, and code in dynamic environments.

OptimusKG is a multimodal biomedical knowledge graph unifying diverse data sources with schema constraints, validated by a multimodal agent to capture knowledge that may precede scientific literature synthesis.

A model-agnostic framework using MCMC to preserve temporal dynamics in synthetic time series generation for improved forecasting.

An unsupervised machine learning workflow for electrofacies classification and porosity characterization using wireline logs in offshore basins with scarce core data.

A conceptual framework for agents to identify and leverage knowledge sources in hybrid human-AI environments.

A tensor learning framework for constructing statistical channel fingerprints in massive MIMO communication systems.

Novel splitting techniques for argumentation formalisms that incorporate collective attacks and supports, generalizing existing frameworks.

This paper introduces a new equilibrium concept in game theory that guarantees existence by minimizing coalitional deviation incentives, with algorithms for average and maximum gain objectives.

This framework proposes personalized digital twins for cognitive decline assessment using multimodal data and uncertainty awareness, showing feasibility in preliminary studies.

This paper investigates how adversarial instructions cause LLMs to rely on positional shortcuts rather than content engagement, revealing vulnerabilities in evaluation methods.

Critiques AI sign language translation tools for perpetuating ableism and standardizing language for profit, arguing for a degrowth approach.

This paper explores design challenges and emerging methodologies for ensuring dependability in autonomous systems integrating AI/ML components.

A conceptual framework to help leaders understand and manage the evolving spectrum of human-AI relationships in decision-making teams.

A framework for theoretically grounded interpretation of AI performances, questioning claims of human-like cognition in large language models.

Investigating splitting on the knowledge base rather than graph instantiation for computationally intractable Assumption-Based Argumentation frameworks.

Investigating how network architecture, task similarity, and representational dimensionality jointly shape learning in continual learning systems.

A framework mapping classroom interaction research dimensions to guide research and tool design in the age of AI.

A review of fairness notions and metrics for distribution network planning and operation, aiming to support transparent decision-making in resource allocation.

Investigating Model Predictive Control for budget allocation under non-stationary returns, showing benefits only when return efficiency has predictable structure.

This paper investigates 'serialization friction' in LLMs, showing that representing 2D structured tasks as 1D sequences hinders performance compared to vision-augmented pathways that preserve spatial layout.

A novel parameterization for Normalized Transformers (νGPT) that enables learning rate transfer across model dimensions and token horizons.

This framework helps educational institutions navigate the adoption of agentic AI by balancing implementation feasibility, adaptation speed, and mission alignment.

This paper proposes a framework for systematically tuning exit strategies in autonomous trading agents to improve risk-adjusted performance.

Investigating why LLM agents deviate from Nash equilibria by analyzing internal mechanisms and demonstrating causal control over strategic behavior.

A theoretical framework explores the geometric and information-theoretic principles of encoder-decoder learning using the Information Bottleneck principle.

Developing a gradient descent-based, physics-constrained Jacobian version of an FCM with residual memory and backpropagation through time.

This research explores how freelance knowledge workers use generative AI for upskilling and identifies challenges in validating acquired skills, offering design recommendations for learning tools.

This paper models threats in LLM-enabled robotic systems, analyzing how conventional, adversarial, and conversational threats interact across architectural boundaries to cause unsafe physical actuation.

Analyzing student-AI interactions in programming education to understand help-seeking behaviors and inform the design of more effective AI teammates.

Exploring interval orders and biorders for rational belief revision, introducing new operators and addressing consistency issues.