Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Benchmarking the Limits of In-Context Reinforcement Learning for Ad-Hoc Teamwork
Yuheng Jing, Kai Li, Jiajun Zhang, Zeyao Ma, Jiaxi Yang, Lei Zhang, Zhe Wu, Jinmin He, Junliang Xing, Jian Cheng
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 |
|---|---|---|---|
| use AD/DPT in-context RL instead of random policy | AD, DPT · ICRL4AHT Track 1 | average return: DPT 12.4 vs Random 5.5; AD 9.1 | Mixed |
| use AD/DPT instead of random policy on unseen layouts | AD, DPT · ICRL4AHT Track 2 | average return: AD 4.4, DPT 13.2 vs Random 11.7 (asymm adv right) | Hurt |
| add teammate actions as input (+TA) (ablation) | AD, DPT · ICRL4AHT Track 1/2 | average return: DPT test wide H1 -23.4 -> -14.5; demo wide H2 9.4 -> 6.8 | Mixed |
| increase Transformer context length from 250 to 2000 (ablation) | AD, DPT · test wide, demo wide | average return | No clear effect |
| scale model size and training budget (Small to Large) (ablation) | AD, DPT · ICRL4AHT Tracks 1 and 2 | mean return: modest gains on teammate track, limited on layout | Mixed |
| replace Transformer with CNN+GRU recurrent tracker (Hybrid-AD) (ablation) | AD · ICRL4AHT | mean return: test wide -7.9 -> -6.5; cramped down 0.8 -> 0.5 | No clear effect |
| use offline RL objective (AMAGO-Offline) instead of action prediction (ablation) | AD/DPT · ICRL4AHT Teammate track | mean return: 8.3 vs AD 8.1, DPT 7.4 | No clear effect |
| extend context to 10,000 (ablation) | AD, DPT · ICRL4AHT selected settings | return | Mixed |
Closest papers by meaning
- Ad Hoc Teamwork via Offline Goal-Based Decision Transformers (ICML 2025 · similarity 0.89)
- ROTATE: Regret-driven Open-ended Training for Ad Hoc Teamwork (NeurIPS 2026 · similarity 0.89)
- N-agent Ad Hoc Teamwork (NeurIPS 2024 · similarity 0.89)
- NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming (NeurIPS 2026 · similarity 0.89)
- Open Ad Hoc Teamwork with Cooperative Game Theory (ICML 2024 · similarity 0.88)
- Minimum Coverage Sets for Training Robust Ad Hoc Teamwork Agents (AAAI 2024 · 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.
