Oxford paper: LLM agents fail zero-shot coordination by treating their own conventions as obvious

rohanpaul_ai · x · 2026-11-08

An Oxford BOLD paper (arXiv:2611.06409, NeurIPS 2026 Poster) studies zero-shot coordination among independently developed LLM agents.

Setup: Using minimal general-purpose scaffolds and 9 small cooperative games with text and code agents built on models like GPT-5.2 and Gemini 3 Pro, the authors isolate coordination failures that complex environments like Hanabi previously obscured. In Cat & Dog, a safe move guarantees 5 points while a signaling convention pays 10 only if both agents match.

Findings: Cross-model agent pairs score below the safe 5 — each agent treats its own arbitrary convention as "the obvious choice," shows weak reasoning about its partner's beliefs, and fails to exploit semantic information as a coordination axis.

Fix: Coordination-friendly definitions (CFDs) — short shared statements of coordination principles — reliably push agents to the safe move, and CFDs can be discovered automatically.

Practical takeaway: for agents that must work with agents you didn't build, instruct them to avoid unstated conventions and prefer actions any partner reads correctly. Code is open-sourced.

Original post →

Oxford paper: LLM agents fail zero-shot coordination by treating their own conventions as obvious

More from coding & agent