Skip to content
LessWrong· Joey Yudelson·Indexed by this site · Originally published· Alignment →· Its place in the scan record →

Research notes: Detecting research sabotage with debate and Slopvestigator

[Research Note] Detecting Research Sabotage with Debate & Slopvestigators

AI summary

The authors tested using debate to detect covert sabotage in AI safety research codebases on Auditing Sabotage Bench, and found that two-round debate (both debaters and judge were GPT-6 Luna) did not outperform a single judge, with an AUROC difference of -0.03 on the original honest control. Debate made the judge more suspicious of all codebases, with true positive rate and false positive rate rising together, and after the authors adjusted the number of rounds, debater effort level and prompts, the results remained within the noise range. The authors judge that the main bottleneck is in hypothesis generation, and turned to testing parallel investigation agents; the capture rate at the extraction stage was 87.9% (10 agents) and 97.0% (20 agents), higher than the 69.7% of a single extractor, but the end-to-end AUROC was still close to the single judge's 0.537. The highest was the variant in which each investigation agent scored itself, with AUROC of 0.742 (fixed honest control) and 0.626 (original honest control); the authors call this a preliminary result. The T v U debate in the appendix shows that introducing an untrusted debater set to mislead did not change debate performance.

Read originallesswrong.com