Reproduce “ToolMisuseBench: An Offline Deterministic Benchmark for Tool Misuse and Recovery in Agentic Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Akshey Sigdel, Rista Baral
This work introduces ToolMisuseBench, an offline and deterministic benchmark designed to evaluate the operational reliability of tool-using agentic systems under explicit budget constraints. It features deterministic fault injection and recovery metrics, revealing that current repair strategies struggle with persistent failures like authorization issues or continuous rate limiting.
Key insights
- The benchmark allows for decomposition of failures to pinpoint their origin (invalid call construction, policy violation, or recovery failure) and offers controlled adjustment of task difficulty.
- Performance is measured using "budgeted success," quantifying the proportion of tasks completed successfully within a tool call cap, emphasizing efficiency-aware performance.
- Deterministic replayability, achieved through task-seeded random streams and declarative trigger definitions, supports iterative research and robust comparison of agent methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.