Controlling the Researcher: AI Control Evaluations for Automated AI R&D

good idea well chosen. interesting framing though « orthogonal » does some heavy lifting here, could be simpler like « embedded / independent » or something. Methodologically, that the side tasks aren’t attempted makes it difficult to test the monitor, and then when it’s fooled to induce bad behavior, to know how reliable that is. Would be good on a larger sample to have an evasion metric to track this. Cool finding in the SFT warmup that the monitor doesn’t catch the misalignment unless explicitly tasked to look at the data -  feels like a common LLM failure mode; not looking at the ground truth, and AFAIK there are scaffolding and adaptive prompting mechanisms in other control evals to account for this. CUDA task is very well done, seems quite novel, and presentation is great overall. Could do with some further stats on these results, and further discussions on seperation of « can current models be made to do this » from « if they did, would we catch it ».

While this paper doesn't introduce any novel insights, it is a solid benchmarking of under represented adversarial setting. It introduces three plausible AI R&D control settings and draws a useful distinction between orthogonal side tasks and sabotage embedded in the primary artifact or pipeline. While results are early, given limited amount of runs, simple red team strategies, etc., the main empirical pattern is believable, even if it doesn't surprise me much. I like simple evals like this, especially if they can later be reused in some more systematic meta review research.