Evolutionary Emergent Misalignment
An evolutionary game-theory benchmark that probes how emergent misalignment changes LLM cooperation, strategy choice, and social behavior in social-dilemma games.
Evolutionary Emergent Misalignment studies emergently misaligned language models through game-theoretic play rather than only helpfulness/harmfulness evaluations.
Building on emergent-misalignment model organisms, the project runs cooperation games (for example Snowdrift and Stag Hunt) against aligned and misaligned checkpoints to measure cooperation rates, antisocial actions, payoffs, and equilibrium tendencies—and how sensitive those behaviors are to prompt or game-setup changes.
Goals
- Compare strategic play before vs. after emergent-misalignment fine-tuning.
- Quantify cooperation vs. utility maximization with repeated and network extensions.
- Provide a finer-grained behavioral assay of the “misaligned persona.”