Research map / Multi-agent systems and games / Games and equilibria
No-Press Diplomacy: Modeling Multi-Agent Gameplay
Philip Paquette, Yuchen Lu, SETON STEVEN BOCCO, Max Smith, Satya O.-G., Jonathan K. Kummerfeld, Joelle Pineau, Satinder Singh, Aaron C. Courville
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 |
|---|---|---|---|
| train DipNet by supervised learning on human trajectories | DipNet · No Press Diplomacy, 1-vs-6 vs Albert 6.0 | TrueSkill: 28.1 - 24.5 | Improved |
| initialize from supervised learning and continue training through self-play | DipNet · No Press Diplomacy, 1-vs-6 vs SL DipNet | win rate: 6.2% vs 14.0% | Mixed |
| add FiLM conditioning (ablation) | DipNet · Human orders test set | unit-order accuracy, greedy: 47.0% -> 47.5% | Improved |
| include previous-phase orders alongside board state (ablation) | DipNet · Human orders test set | unit-order accuracy, greedy: 45.6% -> 47.5% | Improved |
| use location-based attention instead of the average embedding (ablation) | DipNet · Human orders test set | unit-order accuracy, greedy: 46.2% -> 47.5% | Improved |
| retain board information in the decoder (ablation) | DipNet · Human orders test set | unit-order accuracy, greedy: 26.5% -> 47.5% | Improved |
| use the full architecture rather than board-state-only input (ablation) | DipNet · Support-order prediction, 16th location | support accuracy: 25.9% -> 32.2% | Improved |
| use the full architecture rather than removing FiLM (ablation) | DipNet · Support-order prediction, 16th location | support accuracy: 30.3% -> 32.2% | Improved |
| use the full architecture rather than average embeddings (ablation) | DipNet · Support-order prediction, 16th location | support accuracy: 27.9% -> 32.2% | Improved |
| train with self-play rather than stop after supervised learning | DipNet · Agent games | effective cross-power support ratio: 10.2% -> 5.3% | Hurt |
| use board information rather than remove it (ablation) | DipNet · Agent games | effective cross-power support ratio: 0.62% -> 10.2% | Improved |
| use 16 graph convolution layers instead of 8 (ablation) | DipNet · Human orders test set | unit-order accuracy, greedy: 47.4% -> 47.5% | No clear effect |
Closest papers by meaning
- No-Press Diplomacy from Scratch (NeurIPS 2021 · similarity 0.93)
- Mastering the Game of No-Press Diplomacy via Human-Regularized Reinforcement Learning and Planning (ICLR 2023 · similarity 0.93)
- Human-Level Performance in No-Press Diplomacy via Equilibrium Search (ICLR 2021 · similarity 0.92)
- DipLLM: Fine-Tuning LLM for Strategic Decision-making in Diplomacy (ICML 2025 · similarity 0.90)
- Mimicking To Dominate: Imitation Learning Strategies for Success in Multiagent Games (NeurIPS 2024 · similarity 0.86)
- Explaining Decisions of Agents in Mixed-Motive Games (AAAI 2025 · 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.
