When we simulate humans with LLMs in strategic interactions, the efficacy usually depends on the accuracy of simulated beliefs and decisions.
We use cognitive models as executable specifications for LLM agents to capture diverse patterns in human decision-making.
🧵📄👇
No Digg Deeper questions have been answered for this story yet.
🚨New paper! 🚨 You've heard of persona prompting for simulating people, but now you can also use Equation-to-Behavior RL/prompting to supplement it!
Implications for simulations, alignment evals, and RL training robustness!👇👇👇
When we simulate humans with LLMs in strategic interactions, the efficacy usually depends on the accuracy of simulated beliefs and decisions.
We use cognitive models as executable specifications for LLM agents to capture diverse patterns in human decision-making.
🧵📄👇
As LLMs become more capable in strategic interactions, we need more realistic environments that can accurately model human decision-making. Cognitive models offer a principled path toward building these environments!
We also train Sender agents against different kinds of Receivers. Training only against Bayesian receivers is not enough. Training against a mixed population of cognitive models improves average belief change by 2.5–12%.
LLMs are increasingly used to simulate humans in different strategic interactions.
But real humans don’t all reason the same way: some behave roughly like Bayesians, while others exhibit non-Bayesian patterns that persona prompting may hardly capture.
Cognitive science has long formalized these differences as different mathematical models.
We propose Equation-to-Behavior Prompting: using these models as executable specifications to teach LLMs to update beliefs and make decisions.
We test this in persuasion games built from historical legal trials in the Old Bailey. We instantiate the prosecution as the Sender and the court as the Receiver: the Sender presents evidence, while the Receiver updates beliefs and chooses whether to convict or acquit.
We find that large models can often approximate equation-specified belief updating with prompting alone. But smaller models struggle to reliably implement these cognitive rules.
RL substantially improves small models’ ability to follow cognitive models. In out-of-distribution parameter settings, fine-tuning reduces belief error by up to 26.5%.
We introduce Equation-to-Behavior RL: use the cognitive model itself as an executable specification, then reward the model for producing beliefs close to the equation-defined posterior.
Webpage: http://princeton-polaris-lab.github.io/persuasion-webpage Code: http://github.com/chengzr01/persuasion-gym Paper: http://arxiv.org/abs/2606.17657
Work done with amazing co-leader @Zeyu_Shen_yo, and amazing advisors Tom Griffiths @cocosci_lab, and @PeterHndrsn!