Research map / Probabilistic and scientific modeling / Molecules and proteins
What Does Flow-Matching Bring to TD-Learning?
Bhavya Kumar Agrawalla, Michal Nauman, Aviral Kumar
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 distributional backup instead of expected-value backup (ablation) | floq · OGBench hmmaze-large | success rate: 52 -> 30 | Hurt |
| use distributional backup instead of expected-value backup (ablation) | floq · OGBench antmaze-giant | success rate: 86 -> 74 | Hurt |
| use distributional backup instead of expected-value backup (ablation) | floq · OGBench cube-double and hmmaze-medium | success rate: 72 -> 72; 94 -> 94 | No clear effect |
| stale velocity field for first 25% of integration (ablation) | floq · OGBench antsoccer-arena | success rate: 45 -> 50 | Improved |
| stale velocity field for first 25% of integration (ablation) | floq · OGBench cube-double | success rate: 72 -> 84 | Improved |
| stale velocity field for first 25% of integration (ablation) | floq · OGBench hmmaze-medium | success rate: 98 -> 43 | Hurt |
| stale velocity field for all integration steps (100%) (ablation) | floq · OGBench antsoccer-arena | success rate: 45 -> 35 | Hurt |
| freeze early layers of monolithic critic (ablation) | FQL monolithic critic · OGBench | success rate | Hurt |
| add i.i.d. noise to TD velocity targets, flow vs monolithic critic (ablation) | floq · OGBench | success rate | Improved |
| freeze early features during TD learning (ablation) | floq · OGBench | success rate | No clear effect |
| train at a single integration step only (zero only) (ablation) | floq · OGBench | stability | Hurt |
| supervise TD target directly instead of velocity at each step (ablation) | floq · OGBench | success rate | Hurt |
Closest papers by meaning
- Quantile-Coupled Flow Matching for Distributional Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- Composite Flow Matching for Reinforcement Learning with Shifted-Dynamics Data (NeurIPS 2025 · similarity 0.89)
- Flow Q-Learning (ICML 2025 · similarity 0.88)
- The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL. (NeurIPS 2026 · similarity 0.87)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.87)
- Flow Matching for Offline Reinforcement Learning with Discrete Actions (NeurIPS 2026 · 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.
