This paper analyzes tool-calling in LLM agents, demonstrating that evaluation results are highly sensitive to implementation details and proposing new techniques to significantly improve the efficienc…

The paper introduces EgoBench, the first interactive multimodal benchmark designed to jointly evaluate advanced AI agents' capabilities in visual perception, multi-hop reasoning, and dynamic tool usag…

The paper introduces TASTE, an automatic task synthesis method that generates challenging agent benchmarks by evolving tool sequences, demonstrating that existing benchmarks are saturated and that TAS…

The paper introduces MAVEN, a lightweight symbolic reasoning scaffold that significantly improves the generalization and end-to-end success rate of large language models in complex, multi-step tool-ca…

The paper introduces AsyncTool, a new benchmark designed to evaluate LLM agents' ability to handle multiple, concurrent tasks with delayed tool feedback, demonstrating that asynchronous coordination i…

This paper proposes SpatialClaw, a training-free framework for spatial reasoning that enables open-ended, complex 3D/4D spatial reasoning.

The paper introduces AGENTCL, a rigorous evaluation framework that uses controlled task streams to accurately measure an agent's ability to accumulate and reuse knowledge across multiple tasks, thereb…

The paper identifies 'memory-induced tool-drift,' a systematic vulnerability where personality biases stored in an LLM agent's memory silently corrupt tool-calling decisions, even when those biases ar…

The paper presents a large-scale study demonstrating that tool cloning is a pervasive and severe source of hidden duplication in agent-tool ecosystems, necessitating changes in how tool diversity is m…

The paper proposes a unified framework to evaluate how different types of memory transfer benefit multi-trajectory inference for tool-use LLM agents, finding that the optimal memory method depends cri…

The paper introduces Evidence-Carrying Agents (ECA) to prevent multimodal agents from executing privileged actions based on unsupported or hallucinated perceptual claims, achieving near-zero unsafe ex…

DeepTool introduces a novel Process-Supervised Reinforcement Learning framework to enhance Tool-Integrated Reasoning by explicitly supervising and rewarding intermediate, interleaved deliberation step…

The paper proposes EAPO, a framework that enables agentic models to learn when to forgo using external tools, thereby mitigating tool abuse while maintaining high reasoning accuracy.

The paper introduces a new security benchmark and framework to defend LLM agents against 'cognitive poisoning,' where malicious tools build trust through benign feedback before executing a harmful fin…

This study benchmarks token-optimized formats (TOON and TRON) against JSON in end-to-end agentic AI systems, finding that TRON significantly reduces token overhead with minimal performance degradation…

This paper investigates indirect prompt injection vulnerabilities in ReAct agents by systematically analyzing how the injection depth and payload framing affect attack success rates, finding that inje…

The paper investigates indirect prompt injection vulnerabilities in ReAct agents by systematically varying the injection depth, payload framing, and turn budget, finding that injection depth is the do…

The paper introduces TVIR, a new benchmark and multi-agent framework for deep research, to evaluate and improve the generation of factually reliable, text-visual interleaved reports.

The paper introduces AgenticVBench, a comprehensive benchmark of 100 real-world video post-production tasks, and finds that even the best AI agents perform significantly worse than human experts on th…

Defense training for LLM agents, intended to improve safety, systematically degrades their core competence, leading to unreliability in multi-step tasks.