GENEGO: A Disciplined LLM Negotiation Agent with Pairwise Preference Inference

Large language models (LLMs) are fluent conversationalists but remain unreliable negotiators. Single-prompt agents tend to forget past offers, contradict themselves between turns, and concede value just to keep the conversation moving. We present GENEGO, a structured negotiation agent built around three components: a ledger that records every offer the agent has made or agreed to, a partner model that estimates the opponent's preference ordering from the dialog using pairwise A/B comparisons aggregated with a Bradley-Terry model, and a tit-for-tat rule that keeps the agent above its own walk-away threshold. On a multi-issue negotiation benchmark, GENEGO averages 22.0 points per battle against a strong prompt-only baseline's 19.1 (paired t-test, p<0.001). The partner model recovers the exact preference ordering on 82% of turns and the top-priority item on 97%; its bootstrap-based confidence is well calibrated, with accuracy rising from 14% in the low-confidence band to 94% in the high-confidence one. Reliable LLM negotiation needs strategic discipline that the agent follows, a fixed record of every offer made, and a robust way to infer the partner's preferences — GENEGO supplies all three.