Tag
A new 4B parameter AI model trained with Social Reinforcement Learning (SocialRL) outperforms GPT-5 models in negotiation tasks, suggesting that traits making an AI pleasant assistant (agreeableness, transparency) hinder negotiation effectiveness.