Behavioral Fidelity of LLMs in Complex Decision‑Making Environments

Determine how well large language models capture human behavior in complex decision‑making environments characterized by strategic interdependence and endogenous belief formation.

Background

The paper distinguishes between simple decision tasks, where LLMs often align with human responses, and complex multi-agent environments that require anticipating others’ actions and forming beliefs from observed behavior.

The authors highlight that, despite progress, it is still uncertain whether LLMs reproduce human behavior in such strategically interdependent, belief-driven settings, motivating their two-stage framework and empirical tests.

References

It remains unclear how well LLMs capture human behaviors in such settings.

— Improving Behavioral Alignment in LLM Social Simulations via Context Formation and Navigation  (2601.01546 - Kong et al., 4 Jan 2026) in Section 2.2, LLM Social Simulations

In our view, there is no general answer about which theory is right. Whether players hold statistically correct beliefs about opponents' play likely depends on context and experience, both in the laboratory and in the real world.

— Cursed Rationalizability  (2609.28392 - Cohen et al., 23 Sep 2026) in Conclusion, paragraph beginning “These experiments do not distinguish cursed rationalizability from cursed equilibrium.”