AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Weak-to-Strong Generalization with Failure Trajectories

Ruimeng Ye, Zihan Wang, Yang Xiao, Zinan Ling, Manling Li, Bo Hui

ICLR 2026 · Reinforcement learning › Visual and representation RL · approach: RL with tree search for discovery

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 tree-derived preference pairs instead of unstructured random pairs for DPO (ablation)W2SG DPO, Llama3-8B · AlfWorldaverage reward: 60.4 -> 61.9Improved
add MCTS over the trajectory tree instead of Tree DPOW2SG, Llama-2-13B strong · WebShopaverage reward: 53.2 -> 56.9Improved
use weak-model trajectory MCTS instead of plain SFT of the strong modelLlama-2-13B · ScienceWorldaverage reward: 53.6 -> 58.2Improved
raise DPO beta from 0.1 to 0.5 (ablation)W2SG Tree DPO (WTS) · ScienceWorldaverage reward: 54.9 -> 49.2Hurt
raise MCTS tree breadth from 6 to 7 (ablation)W2SG MCTS · ScienceWorldaverage reward: 58.2 -> 54.9Hurt
raise MCTS tree breadth from 5 to 6 (ablation)W2SG MCTS · ScienceWorldaverage reward: 57.6 -> 58.2Improved
increase number of weak trajectories beyond optimal range (ablation)W2SG · AlfWorldaverage rewardMixed
use Llama2-13B instead of Llama2-7B as weak model (ablation)W2SG MCTS, Llama3-8B strong · WebShopaverage reward: 63.7 -> 65.3Improved
use Llama2-13B instead of Llama2-7B as weak model (ablation)W2SG Tree DPO, Llama3-8B strong · WebShopaverage reward: 61.3 -> 62.0Improved
apply W2SG MCTS to Qwen2.5-14B strong with Qwen2.5-7B weakQwen2.5-14B · WebShopaverage reward: 71.7 -> 76.1Improved
apply W2SG MCTS to Llama2-70B strongLlama2-70B · ScienceWorldaverage reward: 64.6 -> 69.1Improved
add best-of-8 sampling to SFT strong modelLlama-2-13B · WebShopaverage reward: 51.0 -> 52.3Improved

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