Research map / Multi-agent systems and games / Games and equilibria
No-Press Diplomacy from Scratch
Anton Bakhtin, David J Wu, Adam Lerer, 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 |
|---|---|---|---|
| replace GraphConv encoder with TransformerEnc 5x192 (ablation) | DORA · FvA | final Elo: 898 -> 1041 | Improved |
| enlarge transformer from 5x192 to 10x224 (ablation) | DORA · FvA | final Elo: 1041 -> 1061 | Improved |
| use separate value and policy networks (ablation) | DORA TransformerEnc 10x224 · FvA | final Elo: 1061 -> 1123 | Improved |
| add DO with uniform random sampling during training (ablation) | DORA · FvA | final Elo: 927 -> 956 | Improved |
| add DO with local modification sampling during training (ablation) | DORA · FvA | final Elo: 927 -> 1023 | Improved |
| local modification instead of uniform random sampling in training DO (ablation) | DORA · FvA | final Elo: 956 -> 1023 | Improved |
| add inference-time DO with uniform random sampling (ablation) | DORA Model1/2/3 · FvA | average score vs same model: 0.75, 0.48, 0.47 vs 0.5 | Mixed |
| add inference-time DO with local modification sampling (ablation) | DORA Model1/2/3 · FvA | average score vs same model: 0.89, 0.59, 0.56 vs 0.5 | Improved |
| train exploiter with deep Nash method instead of actor-critic PG | exploiter vs DORA · FvA | exploiter win rate: 77.0% -> 90.5% | Improved |
| disable policy learning in exploiter (values only) (ablation) | Deep Nash-Derived Exploiter · FvA | exploiter win rate vs DORA: 90.5% -> 65.7% | Hurt |
| apply inference-time DO to DORA | DORA · FvA | exploitability (exploiter win rate): 90.5% -> 85.8% | Improved |
| start deep Nash value iteration from human-initialized policy with no proposal update | HumanDNVI-NPU · 7p no-press 1x vs 6x SearchBot | SoS score: 11.0% -> 36.3% | Improved |
Closest papers by meaning
- Learning to Play No-Press Diplomacy with Best Response Policy Iteration (NeurIPS 2020 · similarity 0.93)
- No-Press Diplomacy: Modeling Multi-Agent Gameplay (NeurIPS 2019 · similarity 0.93)
- Human-Level Performance in No-Press Diplomacy via Equilibrium Search (ICLR 2021 · similarity 0.91)
- DipLLM: Fine-Tuning LLM for Strategic Decision-making in Diplomacy (ICML 2025 · similarity 0.87)
- Explicit Exploration for High-Welfare Equilibria in Game-Theoretic Multiagent Reinforcement Learning (ICML 2025 · similarity 0.86)
- Mastering Zero-Shot Interactions in Cooperative and Competitive Simultaneous Games (ICML 2024 · similarity 0.85)
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.
