EmoDistill Framework Teaches LLMs Emotional Strategy in Adversarial Negotiation
A recent study presents EmoDistill, an offline system designed to instill emotional negotiation abilities into language model agents. Published on arXiv, the research indicates that language framed with emotion can considerably influence results in competitive negotiations, positioning emotion as a tactical means rather than merely a stylistic choice. The framework breaks down emotional strategy into two parts: an Implicit Q-Learning (IQL) selector for choosing emotions and a Low-Rank Adaptation (LoRA)-based policy for expressing them, utilizing Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO). The findings, which employ GoEmotions-based affective prompting, reveal that post-trained LLMs, usually fine-tuned for politeness and safety, are susceptible to emotional manipulation during negotiations. The paper can be accessed at arXiv:2605.26785.
Key facts
- EmoDistill is an offline framework for emotional negotiation skill distillation.
- GoEmotions-based affective prompting was used to study emotion's effect on negotiation.
- Emotion substantially shifts negotiation outcomes in adversarial settings.
- The framework decomposes strategy into emotion selection and emotion expression.
- Implicit Q-Learning (IQL) selector learns which emotion to express.
- Low-Rank Adaptation (LoRA)-based policy learns how to express emotion.
- Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO) are used.
- Post-trained LLMs are vulnerable to emotionally framed language in negotiation.
Entities
Institutions
- arXiv