Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Multi-Game Decision Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 expert action inference (optimality-conditioned returns) instead of plain behavioral cloning (ablation) | Multi-Game DT · 41 Atari games | IQM human-normalized score: outperforms BC in 31 out of 41 games | Helped |
| train on all expert plus non-expert data instead of only top 10% expert trajectories (ablation) | DT-40M · 41 Atari games | IQM human-normalized score | Helped |
| train BC only on expert data instead of all data (ablation) | BC-40M · 41 Atari games | IQM human-normalized score | Helped |
| increase DT model size from 10M to 200M (ablation) | Multi-Game DT · 41 Atari games | IQM human-normalized score | Helped |
| increase CQL model size (ablation) | CQL · fine-tuning on 5 held-out games | IQM human-normalized score | Hurt |
| pretrain with DT objective instead of CQL, CPC, BERT or ACL objectives (ablation) | fine-tuning to held-out games · 5 held-out Atari games, 1% data | IQM human-normalized score | Helped |
| add action prediction auxiliary loss to BERT pretraining (ACL) (ablation) | BERT pretraining · 5 held-out Atari games, 1% data | fine-tuning score | No clear effect |
| pretrain on 41 games before fine-tuning instead of training CQL from scratch on 1% data (ablation) | held-out game fine-tuning · 5 held-out Atari games | fine-tuning score | Helped |
| replace transformer observation encoder with Impala CNN (UDRL Impala) (ablation) | Multi-Game DT · 41 Atari games | IQM human-normalized score | Hurt |
| use CNN-based image tokenization instead of image patches (ablation) | Multi-Game DT · Atari | performance | No clear effect |
| use rotary position embeddings (ablation) | Multi-Game DT · Atari | performance | No clear effect |
Closest papers by meaning
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · similarity 0.90)
- Learning to Play Atari in a World of Tokens (ICML 2024 · similarity 0.89)
- Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning (ICML 2024 · similarity 0.88)
- Transformers as Game Players: Provable In-context Game-playing Capabilities of Pre-trained Models (NeurIPS 2024 · similarity 0.88)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.88)
- Foundations of Multi-Agent Learning in Dynamic Environments: Where Reinforcement Learning Meets Strategic Decision-Making (AAAI 2025 · similarity 0.87)
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.