Can Language Models Sandbag Manipulation?

This work is an extension of van der Weij et al’s work on sandbagging. Exploring non-multiple-choice tests is a natural but interesting and well-chosen extension of the sandbagging demonstration work, especially since the Make Me Pay evaluation chosen is not a static benchmark but a dynamic two-agent scenario more akin to many real-world scenarios. The project demonstrates some good iteration (e.g. the generous v ungenerous version of the mark’s prompt). The results are harder to directly compare against van der Weij et al because the set of models used is different (only Llama-3-70b overlaps between the two). Testing on Claude 3 Opus and GPT-4, where van der Weij et al saw the best performance, would make the comparison stronger, but is naturally left to future work in a hackathon.

A great investigation into a novel capability in modern models and sandbagging is important to understand correctly. I'd be curious to see performance on a similar open-ended benchmark but still with binary or four multiple-choice options to make the relationship to multiple-choice benchmarks explicit and simply change one variation of the experiment at a time. Sandbagging on continuous variables seems exceedingly hard for models due to their general tendency to not be very prompt-principled on continuous variables. This might be possible with more CoT (meta-cognition for calculation of outcomes, similar to human thought prior to action) or simply from next-gen models. Overall, an in-depth exploration with good methods given the limited time of the weekend. Great work!