ASMR-Bench: Auditing for Sabotage in ML Research
This benchmark addresses the need for auditing tools to detect subtle sabotage in AI-conducted research, a critical problem for research integrity as AI systems become more autonomous.
ASMR-Bench is a benchmark for detecting sabotage in ML research codebases, where frontier LLMs and LLM-assisted human auditors struggled, with the best performance achieving an AUROC of 0.77 and a top-1 fix rate of 42%.
As AI systems are increasingly used to conduct research autonomously, misaligned systems could introduce subtle flaws that produce misleading results while evading detection. We introduce ASMR-Bench (Auditing for Sabotage in ML Research), a benchmark for evaluating the ability of auditors to detect sabotage in ML research codebases. ASMR-Bench consists of 9 ML research codebases with sabotaged variants that produce qualitatively different experimental results. Each sabotage modifies implementation details, such as hyperparameters, training data, or evaluation code, while preserving the high-level methodology described in the paper. We evaluated frontier LLMs and LLM-assisted human auditors on ASMR-Bench and found that both struggled to reliably detect sabotage: the best performance was an AUROC of 0.77 and a top-1 fix rate of 42%, achieved by Gemini 3.1 Pro. We also tested LLMs as red teamers and found that LLM-generated sabotages were weaker than human-generated ones but still sometimes evaded same-capability LLM auditors. We release ASMR-Bench to support research on monitoring and auditing techniques for AI-conducted research.