Research map / Multi-agent systems and games / Games and equilibria
Human-Level Performance in No-Press Diplomacy via Equilibrium Search
Jonathan Gray, Adam Lerer, Anton Bakhtin, Noam Brown
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add one-step regret-minimization equilibrium search to the human-trained blueprint | Blueprint in no-press Diplomacy · 1v6 vs. Blueprint | SoS score: 20.2% ± 1.3% -> 52.7% ± 1.3% | Improved |
| add one-step regret-minimization equilibrium search | Blueprint in no-press Diplomacy · 1 expert human vs. 6 bots | human SoS score: 22.5% -> 5.7% | Improved |
| add imitation-learning model improvements | DipNet blueprint · 1 expert human vs. 6 bots | human SoS score: 39.1% -> 22.5% | Improved |
| add combined build-order modeling and encoder dropout (ablation) | DipNet imitation model · Policy accuracy | accuracy: 60.5% -> 62.0% | Improved |
| add encoder changes from prior work (ablation) | DipNet imitation model · Policy accuracy | accuracy: 62.0% -> 62.4% | Improved |
| use webdiplomacy training data only (ablation) | Imitation-learning blueprint · Policy accuracy | accuracy: 62.4% -> 61.3% | Hurt |
| add output featurization (ablation) | Blueprint trained on webdiplomacy data · Policy accuracy | accuracy: 61.3% -> 62.0% | Improved |
| add output featurization (ablation) | Blueprint trained on webdiplomacy data · SoS vs. DipNet, temp=0.1 | SoS score: 0.206 -> 0.188 | Hurt |
| add the improved GNN layer (ablation) | Blueprint trained on webdiplomacy data · Policy accuracy | accuracy: 61.3% -> 62.4% | Improved |
| add the merged GNN trunk (ablation) | Blueprint trained on webdiplomacy data · Policy accuracy | accuracy: 61.3% -> 62.9% | Improved |
| average final-iteration RM policies across random seeds instead of using one run's final policy (ablation) | Regret-matching search · Subgame 1 | exploitability: 0.00880 -> 0.00038 | Improved |
| average final-iteration RM policies across random seeds instead of using one run's final policy (ablation) | Regret-matching search · Subgame 2 | exploitability: 0.0580 -> 0.0077 | Improved |
Closest papers by meaning
- DipLLM: Fine-Tuning LLM for Strategic Decision-making in Diplomacy (ICML 2025 · similarity 0.86)
- A Minimaximalist Approach to Reinforcement Learning from Human Feedback (ICML 2024 · similarity 0.84)
- Preference-CFR: Beyond Nash Equilibrium for Better Game Strategies (ICML 2025 · similarity 0.84)
- Learning Not to Regret (AAAI 2024 · similarity 0.84)
- Human-Aligned Chess With a Bit of Search (ICLR 2025 · similarity 0.84)
- Robust and Diverse Multi-Agent Learning via Rational Policy Gradient (NeurIPS 2025 · similarity 0.84)
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.
