Research map / Architectures and efficiency / Pruning and efficient tuning
On the Effect of Pre-training for Transformer in Different Modality on Offline Reinforcement Learning
Shiro Takagi
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 |
|---|---|---|---|
| Pre-train on language (GPT2) instead of random init | Decision Transformer · Hopper-medium | normalized return: 67.7 -> 79.5 | Improved |
| Pre-train on language (GPT2) instead of random init | Decision Transformer · HalfCheetah-medium | normalized return: 48.6 -> 48.4 | No clear effect |
| Pre-train on language (GPT2) instead of random init | Decision Transformer · Walker 2D-medium | normalized return: 71.0 -> 71.2 | No clear effect |
| Pre-train on images (iGPT) instead of random init | Decision Transformer · Hopper-medium | normalized return: 67.7 -> 2.6 | Hurt |
| Pre-train on images (iGPT) instead of random init | Decision Transformer · HalfCheetah-medium | normalized return: 48.6 -> 1.3 | Hurt |
| Pre-train on images (iGPT) instead of random init | Decision Transformer · Walker 2D-medium | normalized return: 71.0 -> 4.3 | Hurt |
| Remove context (K=1) (ablation) | randomly initialized model · Hopper-medium | normalized return: 67.7 -> 0.24 | Hurt |
| Remove context (K=1) for GPT2 (ablation) | GPT2 Decision Transformer · Hopper-medium | normalized return: 79.5 -> 83.73 | Improved |
| Remove context (K=1) for GPT2 (ablation) | GPT2 Decision Transformer · Walker 2D-medium | normalized return: 71.2 -> 69.0 | No clear effect |
| Remove context (K=1) for GPT2 (ablation) | GPT2 Decision Transformer · HalfCheetah-medium | normalized return: 48.4 -> 47.7 | No clear effect |
| Remove context (K=1) for random init (ablation) | randomly initialized model · Walker 2D-medium | normalized return: 71.0 -> 69.1 | No clear effect |
| Replace block 1 of random init with pre-trained GPT2 block, no context (ablation) | randomly initialized model · Hopper-medium | action error and mean return | No clear effect |
Closest papers by meaning
- Pre-training with Synthetic Data Helps Offline Reinforcement Learning (ICLR 2024 · similarity 0.90)
- Transformers as Decision Makers: Provable In-Context Reinforcement Learning via Supervised Pretraining (ICLR 2024 · similarity 0.88)
- Unleashing the Power of Pre-trained Language Models for Offline Reinforcement Learning (ICLR 2024 · similarity 0.88)
- Position: Do pretrained Transformers Learn In-Context by Gradient Descent? (ICML 2024 · similarity 0.86)
- Born a Transformer -- Always a Transformer? On the Effect of Pretraining on Architectural Abilities (NeurIPS 2025 · similarity 0.86)
- Unveiling Markov heads in Pretrained Language Models for Offline Reinforcement Learning (ICML 2025 · similarity 0.86)
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.
