Abstract:We study whether multiple large language models (LLMs) can autonomously improve each other in a negotiation game by playing, reflecting, and criticizing. We are interested in this question because if LLMs were able to improve each other, it would imply the possibility of creating strong AI agents with minimal human intervention. We ask two LLMs to negotiate with each other, playing the roles of a buyer and a seller, respectively. They aim to reach a deal with the buyer targeting a lower price and the seller a higher one. A third language model, playing the critic, provides feedback to a player to improve the player's negotiation strategies. We let the two agents play multiple rounds, using previous negotiation history and AI feedback as in-context demonstrations to improve the model's negotiation strategy iteratively. We use different LLMs (GPT and Claude) for different roles and use the deal price as the evaluation metric. Our experiments reveal multiple intriguing findings: (1) Only a subset of the language models we consider can self-play and improve the deal price from AI feedback, weaker models either do not understand the game's rules or cannot incorporate AI feedback for further improvement. (2) Models' abilities to learn from the feedback differ when playing different roles. For example, it is harder for Claude-instant to improve as the buyer than as the seller. (3) When unrolling the game to multiple rounds, stronger agents can consistently improve their performance by meaningfully using previous experiences and iterative AI feedback, yet have a higher risk of breaking the deal. We hope our work provides insightful initial explorations of having models autonomously improve each other with game playing and AI feedback.

Active Learning for Sound Negotiations

Soundness in negotiations

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

The Complex Negotiation Dialogue Game

Negotiation as concurrency primitive

How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis

Negotiation Games (with abstract)

Improving Language Model Negotiation with Self-Play and In-Context Learning from AI Feedback

Emergent Communication through Negotiation

Indirect Dynamic Negotiation in the Nash Demand Game

Language of Bargaining

Decoupling Strategy and Generation in Negotiation Dialogues

It Takes Two to Negotiate: Modeling Social Exchange in Online Multiplayer Games

Evaluating Language Model Agency through Negotiations

Automated Configuration of Negotiation Strategies

Negotiation Games

Cooperation, Competition, and Maliciousness: LLM-Stakeholders Interactive Negotiation

Targeted Data Acquisition for Evolving Negotiation Agents

Sound Algorithms in Imperfect Information Games

Online Active Learning For Sound Event Detection

Semi-Supervised Active Learning for Sound Classification in Hybrid Learning Environments