prosocial-agents

The paper proposes calibrating an artificial guilt reward signal from human neural and behavioral fMRI data, then embeds it in multi-agent PPO agents. The neurally calibrated agents match human social decision rates far better than hand-tuned or selfish baselines.