PartHackBench for Evaluating Long-Horizon Tool Agents via Partial Credit
September 25, 2026
PartHackBench introduces a controlled methodology to prevent score inflation in tool-agent evaluation by ensuring adversarial trajectories match historical ones in state-predicate satisfaction and agent attribution. In testing, historical credit resulted in a mean score inflation of 0.252.
HOW THIS AFFECTS YOU
●
builderThis helps you more accurately assess whether your agent's progress is genuine or just temporary state changes.
●
researcherYou can use this to mitigate the confound of rewarding temporary milestones in agent benchmarks.