Security architecture for Computer Use Agents using strict isolation between trusted task planning and untrusted environment observations to defend against prompt injection.
MAS-Orchestra framework for improving multi-agent system design through holistic orchestration and controlled benchmarks for enhanced reasoning and coordination.
DFAH framework for ensuring reproducibility and consistency in tool-using LLM agents deployed in financial services through determinism and faithfulness testing.
BioAgent Bench: benchmark suite with end-to-end bioinformatics tasks (RNA-seq, variant calling, metagenomics) for evaluating AI agent performance and robustness.
Framework addressing reward overoptimization in RLHF by moving beyond semantic information to capture true human intent and prevent policy exploitation.
NAAMSE framework for evaluating AI agent security using evolutionary optimization and genetic prompt mutation to simulate adaptive multi-turn adversaries.
Framework and benchmark (AgoraBench) for improving LLM negotiation capabilities through utility-focused feedback, spanning nine complex bargaining scenarios.
Research on combining reinforcement learning with verifiable rewards across multiple domains for LLMs to achieve expert-level performance in coding, math, and other specialized areas.
SkillsBench: benchmark with 86 tasks across 11 domains measuring effectiveness of agent skills with curated and self-generated skill evaluations.
Geometric taxonomy of LLM hallucinations identifying unfaithfulness, confabulation, and factual error with detection methods using semantic grounding.
Investigation of lightweight automated AI pipelines for research-level mathematics using LLMs beyond competition benchmarks to practical applications.
ABD: benchmark for default-exception abduction in finite first-order worlds with SMT verification for sparse exception formula generation.
INDUCTION: benchmark for finite structure concept synthesis in first-order logic with exact model checking verification across three observation regimes.
Multi-KPI benchmark for multi-agent reinforcement learning on energy management tasks using CityLearn environment for smart city optimization.
ARLArena: unified framework for stable agentic reinforcement learning addressing training instability and collapse in multi-step interactive tasks.
Analysis of AI agents with skills for social science research, comparing autonomous agent capabilities to chatbots with multi-step reasoning and tool access.
Analysis decomposing physician disagreement in HealthBench medical AI evaluation dataset to understand variance sources and observable features.
Study of sample-efficient generalized planning using learned transition models, comparing Transformer-based approaches to classical symbolic methods.
Evaluation of multimodal LLM reasoning on ECG signals with focus on verifying validity of clinical reasoning traces beyond proxy metrics.
Extended validation of Explainability Solution Space framework across employee attrition and urban resource allocation domains.
HarmonyCell: end-to-end agent framework using LLMs to handle semantic and distribution shifts in single-cell perturbation modeling across incompatible datasets.
LLM-assisted semantic option discovery for deep reinforcement learning to improve data efficiency, interpretability, and cross-environment transferability.
Framework for benchmarking logical reasoning agents using an assessor agent to enforce budgets, parse outputs, and record structured failure types with reproducible evaluation.
CDD method identifies data contamination in small language models by measuring output distribution peakedness, tested on 70M-410M parameter models across benchmarks.
Jagarin: three-layer architecture for AI agents on mobile devices using structured hibernation and on-device wake networks to balance battery life with responsive agent behavior.
STRUCTUREDAGENT: LLM-based web agent using AND/OR tree planning for long-horizon web tasks with improved memory and reasoning.
OPENDEV: open-source terminal-native coding agent for long-horizon development tasks with scaffolding and context engineering.
RoboLayout extends vision language models for differentiable 3D scene generation feasible for embodied agent interaction.
Neural network-based online algorithms for detecting change points in time series with linear computational complexity.
Overview of automated reinforcement learning automating problem modeling, algorithm selection, and hyperparameter tuning.
Survey of cognitive modeling in robotics and AI agents using utility theory for motivation, emotion, and perception.
Loop-based algorithm for online conflict-free scheduling and routing of automated guided vehicles with arbitrary capacity.
Survey of computerized adaptive testing systems using machine learning for personalized and efficient assessment.
Framework (FEX) for generating efficient model explanations using policy gradient optimization without extensive model queries.
Analysis of corruption phenomena in few-shot fine-tuned diffusion models with Bayesian neural network mitigation.
Optimal transport-based method for training robust neural networks resistant to adversarial attacks.
Variational learning approach for Gaussian Process Latent Variable Models using stochastic gradient annealed importance sampling.
Framework for AI-powered automated test case generation and validation to improve software testing efficiency and coverage.
Analysis of spiking neural networks energy efficiency accounting for data movement and memory access overheads beyond computation.
Visual prompting mechanism (PiVOT) for improving discriminative capability in generic object tracking across video frames.
Method for learning dynamical systems from partially observed data using neural delay differential equations inspired by Mori-Zwanzig formalism.
Online training method for spiking neural networks using hybrid-driven LIF model to address gradient discrepancy and improve efficiency.
Puppet-CNN framework modeling convolutional layer parameters as continuous dynamical system instead of discrete fixed-size layer stack.
Adaptive diffusion approach where generation trajectory adjusts per input based on complexity instead of using fixed denoising process.
Research on whether competing multi-armed bandit agents exhibit collusion in repeated games, testing algorithmic behavior without explicit coordination.
Vision-language model evaluation framework using VLMs for multi-modal, multi-task, multi-criteria assessment of generated outputs.
Method learning symbolic world models from images using pretrained vision-language models for zero-shot generalization in long-horizon robot planning.
Semi-supervised adversarial training approach using latent clustering to reduce data requirements and training time for robust model development.
Ensemble framework for neural machine translation that improves low-resource language pair quality without training multiple full models.
Neural network analysis method using gradients to identify and modify weights encoding social biases like gender, race, and religion.