Research map / Reinforcement learning / Policy gradient and actor-critic
One-Step Generative Policies with Q-Learning: A Reformulation of MeanFlow
Zeyuan Wang, Da Li, Yulin Chen, Ye Shi, Liang Bai, Tianyuan Yu, Yanwei Fu
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 residual reformulation g = a_t - u instead of naive MeanFlow noise-to-velocity-to-action (ablation) | one-step policy with Q-learning · antsoccer-arena-singletask | success rate | Improved |
| increase candidate size K in value-guided rejection sampling from 1 to 5 (ablation) | one-step policy · OGBench tasks (humanoidmaze-large, antmaze-large-play, visual puzzle-3x3) | success rate | Improved |
| increase candidate size K to 20 (ablation) | one-step policy · OGBench tasks | success rate | Hurt |
| vary behaviour-cloning coefficient alpha (even with adaptive tuning) (ablation) | one-step policy · Puzzle3x3, humanoidmaze-large, antmaze-medium-play | success rate | Mixed |
| one-step MeanFlow policy versus FQL on puzzle-3x3 | one-step policy · OGBench puzzle-3x3-singletask | success rate (%): 30 -> 66 | Improved |
| one-step MeanFlow policy versus FQL on cube-double and antmaze-giant (offline) | one-step policy · OGBench cube-double / antmaze-giant singletask | success rate (%): 29 -> 3 (cube-double); 9 -> 0 (antmaze-giant) | Hurt |
| online fine-tuning after offline training | one-step policy · antmaze-giant-navigate, cube-double-play | success rate (%): 0 -> 82 (antmaze-giant); 3 -> 95 (cube-double-play) | Improved |
Closest papers by meaning
- Fast and Highly Expressive Policy Learning for Offline Reinforcement Learning via Bootstrapped Flow Q-Learning (ICML 2026 · similarity 0.93)
- Aligning Flow Map Policies with Optimal $Q$-Guidance (NeurIPS 2026 · similarity 0.91)
- Flow Q-Learning (ICML 2025 · similarity 0.91)
- Flow-Based Policy for Online Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Flow Matching for Offline Reinforcement Learning with Discrete Actions (NeurIPS 2026 · similarity 0.89)
- Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport (NeurIPS 2026 · 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.
