External validity of training versus prompting incentive strategies

Determine which strategy for inducing models to respond to incentives—training models to possess preferences for particular rewards or prompting models to act as though they are incentivized—produces more externally valid conclusions about cooperative behavior.

Background

The paper argues that researchers face a fundamental difficulty when manipulating incentives for AI models: it is unclear what models intrinsically want, and human incentives such as monetary payments may not transfer appropriately to LLMs. The authors identify two possible approaches: prompting models to role-play situations involving in-scenario rewards, or training models to respond to particular incentives, such as possessing a fictitious currency.

The unresolved issue is whether training-based incentive formation or prompt-based incentive elicitation yields findings that generalize more reliably beyond the experimental setting. Resolving this question would improve the design and external validity of empirical studies of AI cooperation.

References

It is an unresolved empirical question which of these strategies (i.e., training vs prompting models to act as if they are incentivized in a particular way) produces more externally valid conclusions.

— Identity, Incentives, and Cooperative Behavior for AI: Reverse-Engineering Multiagent Cooperation  (2610.19188 - Perrault et al., 13 Oct 2026) in Section “Research designs to reverse engineer cooperative behavior,” discussion of experimentally manipulating model payoffs