Research map / Reinforcement learning / Policy gradient and actor-critic
Decoupled Q-Chunking
Qiyang Li, Seohong Park, Sergey Levine
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 |
|---|---|---|---|
| decouple critic chunk size from policy chunk size (DQC) instead of coupled QC | Q-chunking · cube-quadruple-100M | success rate (%): 35 -> 92 | Improved |
| use DQC instead of n-step return backup | NS · cube-quadruple-100M | success rate (%): 27 -> 92 | Improved |
| use DQC instead of n-step return backup | NS · humanoidmaze-giant | success rate (%): 95 -> 92 | No clear effect |
| use DQC instead of SHARSA | SHARSA · puzzle-4x5 | success rate (%): 1 -> 96 | Improved |
| use DQC instead of SHARSA | SHARSA · puzzle-4x6-1B | success rate (%): 64 -> 83 | Hurt |
| add distilled critic for partial chunks (DQC vs QC-NS, ha=5) (ablation) | QC-NS · cube-triple-100M | success rate (%): 51 -> 98 | Improved |
| add distilled critic for partial chunks (DQC vs QC-NS, ha=5) (ablation) | QC-NS · cube-quadruple-100M | success rate (%): 53 -> 92 | Improved |
| add distilled critic for partial chunks (DQC vs QC-NS, ha=5) (ablation) | QC-NS · puzzle-4x6-1B | success rate (%): 95 -> 68 | Hurt |
| add proper decoupling with distilled critic instead of naive partial execution (ablation) | DQC-naive · cube-triple-100M | success rate (%): 27 -> 98 | Improved |
| policy chunk size ha=5 instead of ha=1 at h=25 (ablation) | DQC · cube-quadruple-100M | success rate (%): 45 -> 92 | Improved |
| policy chunk size ha=5 instead of ha=1 at h=25 (ablation) | DQC · humanoidmaze-giant | success rate (%): 92 -> 51 | Hurt |
| add optimism in backup or distillation (kappa not 0.5) (ablation) | DQC · OGBench goal-conditioned tasks | success rate (%) | Improved |
Closest papers by meaning
- Chunk-Guided Q-Learning (ICML 2026 · similarity 0.92)
- Reinforcement Learning with Action Chunking (NeurIPS 2025 · similarity 0.91)
- Sequence Compression Speeds Up Credit Assignment in Reinforcement Learning (ICML 2024 · similarity 0.88)
- SEAR: Sample Efficient Action Chunking Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- Long-Horizon Q-Learning: Accurate Value Learning via n-Step Inequalities (NeurIPS 2026 · similarity 0.87)
- A Perspective of Q-value Estimation on Offline-to-Online Reinforcement Learning (AAAI 2024 · 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.
