Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Self-Explaining Deviations for Coordination
Hengyuan Hu, Samuel Sokota, David J Wu, Anton Bakhtin, Andrei Lupu, Brandon Cui, Jakob Nicolaus Foerster
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 |
|---|---|---|---|
| Search for a joint self-explaining deviation from the no-op blueprint | IMPROVISED^E and IMPROVISED^P · Trampoline-Tiger | optimal strategy discovery: 100% | Improved |
| Search for a joint self-explaining deviation from the no-op blueprint | IMPROVISED^E · Trampoline-Tiger | expected return: 0.1 | Improved |
| Use MAPPO instead of joint SED search | Trampoline-Tiger · Trampoline-Tiger | optimal strategy discovery: 2/20 | Hurt |
| Search for joint deviations by Players 1 and 2 | IMPROVISED^E · Hanabi, 200 finesse-able situations where Player 3 responds correctly | finesses: 15 -> 51 | Improved |
| Choose deviations to maximize probability of improvement rather than expected return | IMPROVISED^P · Hanabi, 200 finesse-able situations where Player 3 responds correctly | finesses: 51 -> 62 | Improved |
| Use joint self-explaining deviations during the full game | IMPROVISED^E · Hanabi, full game, 100 game seeds | finesses: 3 -> 29 | Improved |
| Use test-time deviations at finesse-able situations | IMPROVISED^E · Hanabi, finesse-able situations | average score: 17.52 -> 18.58 | Improved |
| Use test-time deviations at finesse-able situations | IMPROVISED^P · Hanabi, finesse-able situations | average score: 17.52 -> 18.36 | Improved |
| Use test-time deviations during the full game | IMPROVISED^E · Hanabi, full game | average score: 17.8 -> 23.54 | Improved |
Closest papers by meaning
- Explaining Decisions of Agents in Mixed-Motive Games (AAAI 2025 · similarity 0.85)
- Observation Interference in Partially Observable Assistance Games (ICML 2025 · similarity 0.84)
- OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination (ICLR 2025 · similarity 0.83)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.83)
- Blindfolded Experts Generalize Better: Insights from Robotic Manipulation and Videogames (NeurIPS 2025 · similarity 0.82)
- The Attacker in the Mirror: Breaking Self-Consistency in Safety via Anchored Bipolicy Self-Play (NeurIPS 2026 · similarity 0.82)
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.
