Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Zero-Shot Coordination among LLM Agents
Adrian Hayler, Shashank Reddy Chirra, Andrei Lupu, Johannes Forkel, Bidipta Sarkar, Siheng Feng, Jakob Foerster
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies zero-shot coordination between independently developed LLM agents, without shared specialised scaffolding, at test time.
Problem
Zero-shot coordination (ZSC) of independently developed LLM agents at test time; existing LLM coordination work often relies on specialised scaffolding that independent agents are unlikely to share.
Approach
Evaluates LLM agents in ZSC settings without assuming shared specialised scaffolding, contrasting with prior LLM coordination work that depends on such scaffolding. Abstract is truncated, so the specific method is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hand-crafted Coordination Friendly Definition (CFD) to prompt | text agents (Grok 4.1 Fast, GPT-5.2) · nine coordination environments | coordination failure rate: reduced in 6/9 (Grok) and 5/9 (GPT-5.2) environments | Improved |
| add hand-crafted CFD to prompt | text agents · Cat & Dog | joint reward: consistent safe fallback reward of 5 | Improved |
| add hand-crafted CFD to code agents | code agents · SP-XP gap across environments | SP-XP gap | No clear effect |
| add hand-crafted CFD | text agents · ad-hoc conventions | convention formation | Hurt |
| use meta-learned CFD instead of hand-crafted | Grok 4.1 Fast text agent · self-play score | score | Improved |
| use meta-learned CFD trained on code agents with a policy critic | other models and generic code scaffold · XP_LLM | XP score: modest gains | Improved |
| use meta-learned CFD across LLM types | code agents · XP_seed | XP score | No clear effect |
| increase reasoning effort or model size | text agents · Cat & Dog | reveal rate / ZSC performance | Mixed |
| add prompt stating partner is identical instance of the same model | text agents · Cat & Dog | reveal rate: GPT-5.2: ~75% -> ~5% | Hurt |
| relabel signals or translate environment to German | text agents · Cat & Dog | local policy | Mixed |
Limitations
- Only the truncated abstract was available; full methodology and results unknown.
Closest papers by meaning
- More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.89)
- OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination (ICLR 2025 · similarity 0.88)
- LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination (ICML 2025 · similarity 0.88)
- TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (ICML 2026 · similarity 0.88)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
- Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
