Research map / Reinforcement learning / Offline RL
Test-time Offline Reinforcement Learning on Goal-related Experience
Marco Bagatella, Mert Albaba, Jonas Hübotter, Georg Martius, Andreas Krause
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 |
|---|---|---|---|
| add goal-conditioned test-time training on selected data | GC-BC · pointmaze expert | success rate: 0.05 -> 0.78 | Improved |
| add GC-TTT | GC-IQL-DDPG · antmaze play | success rate: 0.24 -> 0.68 | Improved |
| add GC-TTT | SAW · cubesingle and antmaze | success rate: 0.73 -> 0.71 (cubesingle) | No clear effect |
| select random data instead of relevant and optimal data (ablation) | GC-TTT with GC-IQL · pointmaze/antmaze play | success rate: 0.25 -> 0.22 | Hurt |
| add the optimality filter on top of relevance (ablation) | GC-TTT with GC-IQL · pointmaze/antmaze play | success rate: 0.35 -> 0.57 | Improved |
| add the relevance filter on top of optimality (ablation) | GC-TTT with GC-IQL · pointmaze/antmaze play | success rate: 0.34 -> 0.57 | Improved |
| increase TTT frequency (shorter horizon K) (ablation) | GC-TTT with GC-IQL · antmaze play | success rate | Improved |
| scale policy network width to match inference FLOPs instead of using TTT (ablation) | GC-IQL policy · antmaze play | success rate | No clear effect |
| use critic-free variant (trajectory returns) instead of H-step value returns (ablation) | GC-TTT · expert datasets | success rate: pointmaze expert GC-BC: 0.78 -> 0.76 | Mixed |
| increase the number of test-time gradient steps N (ablation) | GC-TTT with GC-IQL · antmaze play | success rate | Mixed |
| replace reward-based relevance with a value-based relevance criterion (ablation) | GC-TTT with GC-IQL · antmaze play | success rate: 0.73 -> 0.68 | No clear effect |
| inject modest Gaussian noise into value estimates (ablation) | GC-TTT with GC-IQL · antmaze play | success rate: 0.73 -> 0.78 (alpha=0.1) | No clear effect |
Closest papers by meaning
- Simple Ingredients for Offline Reinforcement Learning (ICML 2024 · similarity 0.91)
- Option-aware Temporally Abstracted Value for Offline Goal-Conditioned Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- A Clean Slate for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Stitching Sub-trajectories with Conditional Diffusion Model for Goal-Conditioned Offline RL (AAAI 2024 · similarity 0.89)
- Model Selection for Off-policy Evaluation: New Algorithms and Experimental Protocol (NeurIPS 2025 · similarity 0.89)
- Is Value Learning Really the Main Bottleneck in Offline RL? (NeurIPS 2024 · similarity 0.89)
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.
