AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting

Wenhao Zhang, Yuexiang Xie, Yuchang Sun, Yanxi Chen, Guoyin Wang, Yaliang Li, Bolin Ding, Jingren Zhou

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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
replace rigid SFT-then-RL switch with smoothly decaying SFT weight mu (CHORD-mu)SFT+RL on Qwen2.5-7B · AMCaccuracy: 58.4 -> 60.8Improved
add token-wise weighting phi=p(1-p) on SFT loss (CHORD-phi)GRPO pure RL · AIME25accuracy: 8.54 -> 17.2Improved
add CHORD-phi to pure RLGRPO · BFCL Overallaccuracy: 77.1 -> 78.5Improved
decay mu dynamically instead of fixed mu (ablation)CHORD-mu · reward curvesrewardImproved
minimal SFT before RL (SFT-light) vs original modelSFT-light · BFCL Overallaccuracy: 46.2 -> 34.0Hurt
start RL from SFT-light instead of SFT-bestSFT then RL · BFCL Overall vs AMCaccuracy: tool-use 76.1 -> 77.2; AMC 58.4 -> 52.5Mixed
use reward-adaptive mu (threshold 0.5 vs 0.7) (ablation)CHORD-mu · BFCL Overallaccuracy: 78.1 (tau=0.5), 75.6 (tau=0.7)Mixed
replace p(1-p) with focal-loss weight (ablation)CHORD-phi · BFCL Overallaccuracy: 78.5 -> 73.4Hurt
replace p(1-p) with IS clipping (ablation)CHORD-phi · BFCL Overallaccuracy: 78.5 -> 75.9Hurt
replace p(1-p) with entropy-based weights (ablation)CHORD-phi · AIME2024accuracy: 14.0 -> 17.2 (Entropy Top); AMC23 59.7 -> 52.5 (Entropy Norm)Mixed
apply CHORD-phi on weaker 3B policyQwen2.5-3B-Instruct · AMC23accuracy: 36.2 -> 41.9 vs SFT+RLImproved
apply CHORD-phi to MoE modelPhi-mini-MoE-instruct · BFCL Overallaccuracy: 49.5 -> 61.6Improved

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