Control evaluations test whether safeguards stop an AI agent that intentionally tries to cause harm.
Simulated tool responses can predict AI control safety better than separate testbeds
In LinuxArena tests, an LM-based simulator of tool outputs predicted safety metrics with 3.6 percentage point RMSE, outperforming BashArena's 10.1.
Research Lab
Jeffrey Olmo · David Zhang · Tyler Tracy
Redwood Research
Research Digest··2 min read
The authors propose simulated control evaluations (SCEs), which replace a real environment with a language model that generates tool responses.
§