Agentic data science with the Data Agent Kit
cloud.google.com·8h ago
AI Research
arXiv CS.AI·Sep 4·5 min read
EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation
TL;DR
EmoDistill is a framework that distills emotional negotiation skills from large language models (LLMs) into smaller agents, enhancing their performance in adversarial negotiations. It demonstrates improved negotiation outcomes by effectively incorporating emotional cues, outperforming traditional models in various scenarios.
✦ Why It Matters
Consider integrating emotional intelligence into your AI negotiation systems to improve outcomes.
Key Takeaways
- Explore using EmoDistill for training negotiation agents.
- Implement emotional intelligence in AI to enhance negotiation strategies.
- Evaluate the impact of emotional cues on negotiation outcomes.
- Consider domain-specific adaptations for better performance.
How It Works
EmoDistill operates by first using an Implicit Q-Learning (IQL) selector to identify which emotion to express based on the current negotiation state. Then, it employs a 7B parameter policy model that has been fine-tuned through Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO) to articulate the chosen emotion effectively in negotiation utterances.