AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Beyond Policy Training: Recursive Solution Search from Unannotated Videos

Lipeng Wan, Jianhui Gu, Junjie Ma, Anbang Wang, Xuguang Lan

ICML 2026 · Reinforcement learning › Visual and representation RL · approach: Hierarchical skills and priors

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 changedWhereMeasuredEffect
add the 50-sample warm-up set with coarse stage labels (ablation)PFR-Search backward planning · Minecraft (5 tasks average)success rate (%): 24.8 -> 52.0Improved
increase unannotated videos from 300 to 600 to 1200 clips (ablation)PFR-Search backward planning · Minecraft (5 tasks average)success rate (%): 24.0 -> 43.4 -> 52.0Improved
add in-flow consistency (ablation)PFR-Search backward planning · Minecraft (5 tasks average)success rate (%): 50.8 -> 52.0Improved
add cross-flow consistency (ablation)PFR-Search backward planning · Minecraft (5 tasks average)success rate (%): 49.2 -> 52.0Improved
add both in-flow and cross-flow consistency (ablation)PFR-Search backward planning · Minecraft (5 tasks average)success rate (%): 48.8 -> 52.0Improved
add cross-flow consistency (ablation)PFR-Search backward planning · Minecraft Harvest logsuccess rate (%): 63.0 -> 62.0Hurt
add in-flow consistency (ablation)PFR-Search end-to-end search · Minecraft (5 tasks average)success rate (%): 15.0 -> 16.0Improved
add cross-flow consistency (ablation)PFR-Search end-to-end search · Minecraft (5 tasks average)success rate (%): 14.4 -> 16.0Improved
add both consistencies (ablation)PFR-Search end-to-end search · Minecraft (5 tasks average)success rate (%): 14.2 -> 16.0Improved
add task-flow guidance and replay over random explorationrandom exploration · Minecraft (5 tasks average)forward search success rate (%): 1.8 -> 16.0Improved
replace learned task flows with oracle task flows from human demonstrationsPFR-Search end-to-end · Minecraft (5 tasks average)forward search success rate (%): 16.0 -> 27.8Improved
increase the search budgetPFR-Search vs random exploration · Minecraft (two tasks)success rate: PFR-Search improves steadily; random only marginallyMixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy