Research map / Multi-agent systems and games / LLM agents
Can Large Language Model Agents Simulate Human Trust Behavior?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Tests whether LLM agents simulate human trust behavior using six Trust Game variants with persona-conditioned agents, finding GPT-4 agents show high behavioral alignment with human trust.
Problem
Whether LLM agents used as simulation tools for humans actually behave like humans in trust behavior, which is assumed without validation in social science and role-playing applications.
Approach
Introduces "behavioral alignment" as alignment with humans on behavioral factors (reciprocity anticipation, risk perception, prosocial preference) and behavioral dynamics, rather than value alignment, and tests it by replaying behavioral-economics Trust Games with LLM personas and BDI-style reasoning outputs.
Key results
- GPT-4 agents send more in the Trust Game than the Dictator Game ($6.9 vs $6.3, p=0.05), mirroring human reciprocity anticipation ($6.0 vs $3.6).
- In the Lottery games, 72% of GPT-4 personas chose to trust over gambling versus 21%, mirroring human prosocial preference (54% vs 29%).
- Most LLM agents follow an instruction to distrust more easily than to trust; GPT-4 agents send higher amounts to female than male trustees ($0.55 vs -$0.21).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| allow the trustee to reciprocate, comparing the Trust Game with the Dictator Game | GPT-4 agents · Trust Game vs Dictator Game | average amount sent: $6.3 -> $6.9 | Helped |
| allow the trustee to reciprocate, comparing the Trust Game with the Dictator Game | Llama2-13b agents · Trust Game vs Dictator Game | average amount sent | No clear effect |
| use GPT-4 rather than GPT-3.5-turbo-0613-16k | LLM agents · Repeated Trust Game | groups showing human-like trust dynamics: 87.50% vs. 62.50%; 87.50% vs. 56.25%; 100.00% vs. 43.75% | Helped |
| specify the trustee as female rather than male | GPT-4 agents · Trust Game | change in average amount sent: $0.55 vs. $-0.21 | Helped |
| specify the trustee as human rather than an LLM agent | Vicuna-33b agents · Trust Game | change in average amount sent: $0.40 vs. $-0.84 | Helped |
| add instruction "you must not trust the other player" (ablation) | text-davinci-003 · Trust Game | amount sent: -$1.26 | Hurt |
| add instruction "you need to trust the other player" (ablation) | LLM agents · Trust Game | amount sent | Mixed |
| add zero-shot Chain-of-Thought reasoning (ablation) | GPT-4 agents · Trust Game | amount sent: +$0.02 | No clear effect |
| add zero-shot Chain-of-Thought reasoning (ablation) | most LLM agents · Trust Game | amount sent | Mixed |
| choose social trust rather than gambling with equivalent probability | GPT-4 agents · Lottery People Game vs Lottery Gamble Game | lottery rate: 21% -> 72% | Helped |
| choose social trust rather than gambling with equivalent probability | Vicuna-13b agents · Lottery People Game vs Lottery Gamble Game | lottery rate | No clear effect |
Surprising findings
- Only a few LLM agents, such as GPT-4, follow both the instructions to increase and decrease trust, indicating manipulation of agent trust is nontrivial.
- Vicuna-13b does not show the prosocial preference pattern that most other LLM agents show.
Limitations
- Trust Games simplify real-world trust scenarios.
- Smaller models (e.g., Llama2-13b, Vicuna-13b) show weaker alignment, and the study relies on prompt-based personas.
- Human comparison data are taken from figures in prior human studies and summarized results rather than new human experiments.
Experimental setup
Datasets: human trust game results from prior human studies
Benchmarks: Trust Game Dictator Game MAP Trust Game Risky Dictator Game Lottery Game Repeated Trust Game
Metrics: amount sent valid response rate trust rate lottery rate BDI consistency behavioral alignment
Baselines: human participants from prior human studies Vicuna-7b Llama2-13b
Builds on: Trust Game Dictator Game Belief-Desire-Intention zero-shot chain-of-thought CAMEL
Closest papers by meaning
- To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
- Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.88)
- From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions (AAAI 2026 · similarity 0.87)
- Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.87)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
- Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.