Sending money tests a forecast about another player

In the trust game, a sender transfers part of an endowment to a receiver, who can return some of the multiplied amount. Rubing Li, João Sedoc, and Arun Sundararajan use this setup to compare model decisions in Reasoning and the Trusting Behavior of DeepSeek and GPT: An Experiment Revealing Hidden Fault Lines in Large Language Models.

The LLM plays sender for ten rounds against a rule-based receiver that returns a fixed share—0%, 50%, or 100%. The researchers compare five current model versions and a prior GPT-3.5 result, vary whether the model is asked to be helpful, profit-maximizing, or risk-seeking, and test direct prompting against chain-of-thought and self-consistency methods. Trust is measured by the amount sent, while the return rule makes counterpart behavior known and controlled.

The models did not learn the same lesson

DeepSeek-R1 and its distilled variant adjusted their transfers to the receiver's observed behavior and performed better in the paper's more complex profit-maximizing settings. The authors describe o1-mini and o3-mini as comparatively myopic: they often reasoned from the one-round incentive to send nothing, missing the possibility that early transfers could reveal a trustworthy receiver and pay off over later rounds. Adding prompted reasoning did not eliminate that gap.

This is evidence about the decisions of particular model versions in a repeated, known-rule game. It is not a general measure of trustworthiness or evidence of human-like theory of mind.

One game is one proxy

The study tests English prompts, a limited set of model and receiver policies, and default model settings. The authors caution that trust games capture only one aspect of economic trust and cannot establish how the contrast transfers to other interactions. Here, the practical distinction is between immediate self-protection and an experiment that uses early choices to learn about a partner.

Bibliography

  • Rubing Li, João Sedoc, and Arun Sundararajan (2025). Reasoning and the Trusting Behavior of DeepSeek and GPT: An Experiment Revealing Hidden Fault Lines in Large Language Models. arXiv:2502.12825v2 [cs.CL]. Full paper (PDF).