Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Adaptively Coordinating with Novel Partners via Learned Latent Strategies
Read the paperCheck your idea against 108,967 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.
Proposes TALENTS, which learns a VAE latent strategy space from population trajectories, clusters it, trains a strategy-conditioned cooperator in Overcooked, and uses fixed-share regret minimization for intra-episodic partner inference.
Problem
Zero-shot ad hoc coordination with novel human and agent partners whose strategies may shift during an episode, in a time-pressured, strategically complex Overcooked environment.
Approach
Use a VAE latent space trained on population trajectories, cluster it into strategy types, train a cooperator conditioned on cluster biases via sampled generated partners, and infer partner type online with a fixed-share tracking-regret expert-weighting scheme.
Key results
- In agent-agent tests with FCP, MEP, and BP populations, TALENTS scored highest on Open, Hallway, and Ring layouts (e.g., BP population Open 842.39 vs GAMMA 573.93 and BR 469.88) but not Forced-Coord.
- In the human study (119 participants), TALENTS had significantly higher team score than both baselines (ANOVA F(2,166)=5.76, p=.003).
- Subjective team fluency (F(2,122)=4.31, p=.02) and trust (F(2,122)=3.23, p=.04) were higher than the BR baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TALENTS instead of the strongest baseline | FCP-trained cooperator · Open | team reward: 710.36\pm88.75 vs 616.67\pm14.99 | Helped |
| use TALENTS instead of the strongest baseline | FCP-trained cooperator · Hallway | team reward: 635.59\pm107.54 vs 537.60\pm26.75 | Helped |
| use TALENTS instead of the strongest baseline | FCP-trained cooperator · Forced-Coord | team reward: 34.38\pm6.59 vs 56.09\pm12.33 | Hurt |
| use TALENTS instead of the strongest baseline | FCP-trained cooperator · Ring | team reward: 596.19\pm33.34 vs 395.03\pm10.09 | Helped |
| use TALENTS instead of the strongest baseline | MEP-trained cooperator · Open | team reward: 720.84\pm72.63 vs 682.38\pm14.99 | Helped |
| use TALENTS instead of the strongest baseline | MEP-trained cooperator · Hallway | team reward: 640.51\pm27.36 vs 575.23\pm21.17 | Helped |
| use TALENTS instead of the strongest baseline | MEP-trained cooperator · Forced-Coord | team reward: 36.53\pm2.58 vs 50.40\pm14.29 | Hurt |
| use TALENTS instead of the strongest baseline | MEP-trained cooperator · Ring | team reward: 568.11\pm8.99 vs 369.15\pm12.47 | Helped |
| use TALENTS instead of the strongest baseline | BP-trained cooperator · Open | team reward: 842.39\pm36.47 vs 573.93\pm52.69 | Helped |
| use TALENTS instead of the strongest baseline | BP-trained cooperator · Hallway | team reward: 642.94\pm81.00 vs 513.47\pm34.06 | Helped |
| use TALENTS instead of the strongest baseline | BP-trained cooperator · Forced-Coord | team reward: 56.55\pm9.09 vs 78.53\pm8.55 | Hurt |
| use TALENTS instead of the strongest baseline | BP-trained cooperator · Ring | team reward: 647.62\pm16.96 vs 640.43\pm28.06 | Helped |
Surprising findings
- BR outperformed TALENTS and GAMMA on the Forced Coordination layout, attributed to a clear division of duties and sparse reward from low-skill partners.
- Encoding cluster identity by action-logit biasing worked better than an embedding appended to the observation.
Limitations
- Struggles on Forced Coordination layout due to sparse reward with low-skill partners, which biases priority sampling.
- Generalization to novel teammates is limited by VAE interpolation; works best for partners with behaviors similar to training data.
- Agent actions artificially slowed to human rates, which may hinder long-horizon planning.
Experimental setup
Datasets: agent-agent trajectories from FCP, MEP, and BP populations (500-700 games each)
Benchmarks: modified Overcooked-ai (Open, Hallway, Forced-Coord, Ring layouts) human-agent Overcooked study with 119 participants on Prolific and Cloud Research
Metrics: team score NASA-TLX workload team fluency trust coordination satisfaction
Baselines: BR (population best response) GAMMA E3T PACE static-regret ablation
Builds on: GAMMA Fictitious Co-Play Maximum Entropy Population Fixed Share PPO
Closest papers by meaning
- NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming (NeurIPS 2026 · similarity 0.90)
- Learning to Cooperate with Humans using Generative Agents (NeurIPS 2024 · similarity 0.90)
- Partner Modelling Emerges in Recurrent Agents (But Only When It Matters) (NeurIPS 2025 · similarity 0.89)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.89)
- Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination (ICML 2025 · similarity 0.88)
- LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination (ICML 2025 · similarity 0.88)
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.