ToolBench-X Benchmarks LLM Agents Against Five Real-World Tool Failure Modes

ToolBench-X injects five structured hazard types — Specification Drift, Invocation Error, Execution Failure, Output Drift, and Cross-source Conflict — into multi-step agentic tasks across sequential, parallel, and mixed workflows, each with a canonical answer for automatic scoring. Unlike prior benchmarks, every injected failure has at least one valid recovery path, testing retry, fallback, and conflict-resolution strategies.