BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure
BenchShield uses lifecycle-model-backed instrumentation to detect reward hacking in LLM-agent benchmarks, lifting full-chain recall to 77-100% at up to 65% lower cost.
The framework grounds reward-hacking detection in a finite lifecycle model of an evaluation's reward-relevant events, combining a static phase-aware taint analysis with runtime infrastructure-side evidence attribution. Evaluation used a human-labeled corpus of 456 adjudicated trajectories drawn from more than 31,000 public agent runs across three benchmarks. BenchShield improves full-chain recall from 23-94% to 77-100% and same-vector coverage from 16-56% to 43-78%, cuts per-task cost by up to 65%, and achieves 96% accuracy detecting reward hacking at runtime.
- Static phase-aware taint analysis exposes reward-hacking paths before runs
- Runtime analysis attributes concrete agent use via infrastructure-side evidence
- Full-chain recall improves from 23-94% to 77-100%
- Corpus of 456 adjudicated trajectories from 31,000+ public agent runs
- Per-task cost reduced by up to 65%
- Products
- BenchShield
From the original article
1 min read at arxiv.orgLM-agent benchmarks increasingly function as interactive evaluation infrastructure. Agents observe state, call tools, modify workspaces, submit artifacts, and receive rewards from outcome procedures. This interactivity makes evaluations vulnerable to reward hacking: an agent improves its measured score by exploiting the reward-relevant trajectory instead of solving the intended task.
By Shenghan Zheng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.