The Wise Negotiator
The Wise Negotiator is a universal negotiation environment where an LLM agent learns WHO it’s talking to using Bayesian belief updates — and shifts strategy accordingly.
Core trick
Bayesian belief tracker updates buyer-personality probabilities each round and injects a strategy hint into the LLM's observation before GRPO training.
Tags
Description
The Wise Negotiator is a universal negotiation environment where an LLM agent learns WHO it’s talking to using Bayesian belief updates — and shifts strategy accordingly. Built on OpenEnv, it supports multiple domains: guitar sales, car sales, property, and an HR hiring mode where the LLM plays recruiter against three candidate personalities (GrowthSeeker, StabilitySeeker, MarketExplorer). Every round, a belief tracker updates probabilities based on price movement and language signals, injecting a strategy hint directly into the observation. The agent is trained with GRPO using Qwen2.5-1.5B-Instruct, LoRA r=16, num_generations=2 (2-GRPO), and beta=0.0 (DAPO). Reward is composite: 70% price outcome + 20% empathy + 10% efficiency.
README excerpt
# 🎸 The Wise Negotiator ### AI That Reads Who You Are | OpenEnv Hackathon SF 2026 [](https://youtu.be/agAIMZD7BoI) > Before training, the AI immediately caves. After GRPO training it reads who it's talking to and closes deals above its floor. --- ## 🧠 What It Does Most negotiation agents react to **what** the buyer says. The Wise Negotiator figures out **WHO** the buyer is — and shifts strategy accordingly. Every round, a Bayesian belief tracker updates probabilities across buyer/candidate personalities based on price movement size and language signals. That belief state is injected directly into the LLM's observation, conditioning its response on WHO it's facing, not just WHAT was said. --- ## 🏗️ Architecture ``` Opponent Signal → Bayesian Belief Update → Strategy Hint → LLM Response → Reward ``` - **Sell mode** (guitar, car, property): LLM plays seller vs rule-based buyer - **Hire mode** (salary/HR): LLM plays recruiter vs rule-based candidate - **Belief tracker** updates P(personality) each round from price moves + language - **Reward** = 70% price outcome + 20% empathy + 10% efficiency --- ## 👥 Personalities ### Sell Mode — Buyer Personalities | Personality | Behavior | Signal | Counter Strategy | |---|---|---|---| | Haggler | Starts low ($42), creeps up $3/round | Large moves | Hold firm, they come to you | | Anchor | Opens at $44, barely moves $2/round | Tiny moves + "firm" language | Reframe value