Research map / Multi-agent systems and games / LLM agents
Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities
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.
Introduces ProSim, a simulation framework with individual, scenario, interaction (small-world network), and intervention modules to study prosocial behavior, third-party punishment, and inequity contagion in LLM agents versus 104 human participants.
Problem
Existing evaluations of LLM prosociality rely on static, economically framed paradigms and do not capture how prosocial behavior evolves over time or responds to structural policy inequities.
Approach
A modular simulation framework combining demographic and psychological agent profiles, six prosocial scenarios, small-world network interactions, and policy interventions including reward and burden asymmetry, benchmarked against human participants.
Key results
- GPT-4o achieved the strongest correlation with human prosocial ratings across six scenarios (r = 0.955, p = 0.002).
- Under reward asymmetry prosocial scores declined 25-32%, and under burden asymmetry 19-23%.
- Among LLMs, GPT-series models showed the strongest human alignment in third-party punishment, with threshold-based fairness judgments.
Surprising findings
- GPT-3.5 showed higher prosociality (mean 4.875) than humans (4.226), and GPT-series models even exceeded human punishment rates in highly unfair trials.
- Economic incentives were highly effective for humans but produced mixed effects in LLMs.
Limitations
- Human benchmark is small (104 participants) and relies on online recruitment.
- Results rely on self-reported Likert intentions and hypothetical scenarios rather than real behavior.
- Temperature set to 0 for reproducibility, limiting analysis of response variability.
Experimental setup
Datasets: NBS demographic distributions human participant data (104 online participants)
Benchmarks: ProSim six prosocial scenarios third-party punishment task (60 trials)
Metrics: mean 7-point Likert prosocial intention Pearson correlation with human ratings punishment rate SHAP values relative change in prosocial intention perceived unfairness correlation
Baselines: human participants
Builds on: Watts-Strogatz small-world network third-party punishment paradigm Big Five personality
Closest papers by meaning
- Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models (NeurIPS 2026 · similarity 0.88)
- Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (NeurIPS 2024 · similarity 0.86)
- Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.86)
- Can Large Language Model Agents Simulate Human Trust Behavior? (NeurIPS 2024 · similarity 0.86)
- Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View (ICLR 2025 · similarity 0.86)
- From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions (AAAI 2026 · similarity 0.86)
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.