Good concept, fleshed out. Really like the “applied decision making” that combines human-interpretable ideas of ethics with a formalised benchmark based off game theory.

I really like the idea of using LLMs for creating dynamic benchmarks on the fly! I’d be keen to see more data and analysis to get a sense of the strengths and weaknesses of this approach. I do think more analysis of the methodology itself would be needed before worrying too much about the object-level results. Developing this methodological analysis could be a major contribution though.

This idea is super interesting, and I would like to see future work which builds on some of these concepts! As a fan or roleplaying games, I appreciate the stochastic choose-your-own adventure angle, and think that it could be a really good methodology for getting more comprehensive understanding of model behavior in a general sense, not just for this work, which is awesome; I would really like to see this idea developed further.

One aspect that could be valuable to investigate in the near future would be to increase the stochasticity of the setup. More specifically, I think that varying expected payoff of the three choices per question within the game, as supposed to per game itself, could provide interesting insights.  Another direction which may be worth exploring would be to construct scenarios in which an action from any of the three categories was a substantially reasonable course of action, which may require some additional prompting strategies for scenario construction. Finally, your suggested evaluation sounded quite interesting, and I would have liked to see some of its results.

Very interesting, I'd be curious to see more descriptions in the appendix of the concrete scenarios and why you end up seeing 100% cooperation rate with GPT-4. Similarly, we'd get some interesting information out of running this against something like Grok, Claude Sonnet, Llama, and so on, since we'd like to spot models that are not 100% cooperative. It seems some of the content in the repository could have been moved into the report and it would've been interesting to see some type of conversation plots or cooperation rate by scenario effect sizes, but that's of course difficult if all scenarios result in cooperation. In summary, the methodology seems interesting but could be backed up by more results and evaluation on more models.