AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Positive–Unlabeled Reinforcement Learning Distillation for On-Premise Small Models

Zhiqiang Kou, Junyang Chen, Xin-Qiang Cai, Xiaobo Xia, Ming-Kun Xie, Dong-Dong Wu, Biao Liu, Yuheng Jia, Xin Geng, Masashi Sugiyama, Tat-Seng Chua

ICML 2026 · Reinforcement learning › RL from feedback · approach: Theory of RLHF · 0 citations (OpenAlex)

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
use LDL-GRPO instead of SFT onlyLLaMA3-8B-Instruct · PBFGraw win rate vs GPT-4o: 0.482 -> 0.834Improved
replace scalar-reward Anchor-GRPO with label-distribution LDL-GRPO (ablation)LLaMA3-8B-Instruct · PBFGraw win rate: 0.673 -> 0.834Improved
replace Anchor-GRPO with LDL-GRPO (ablation)LLaVA-7B · A-OKVQA-MCraw win rate: 0.090 -> 0.146Improved
use Anchor-GRPO instead of SFT (ablation)LLaVA-7B · A-OKVQA-MCraw win rate: 0.131 -> 0.090Hurt
continue SFT on teacher outputs (SFT->SFT) instead of preference stageQwen2.5-VL-7B / LLaMA3-8B · writing tasksraw win rate: 0.538 -> 0.352 on CFCW (Qwen2.5-VL column)Hurt
induce preferences from one anchor and multiple candidates instead of a single teacher/student pairDPO · six taskswin rateImproved
use LDL-GRPO instead of AnchorRank-DPOLLaMA3-8B-Instruct · PBFGraw win rate: 0.814 -> 0.834Improved
vary sampling temperature from 0.6 to 1.5 (ablation)LDL-GRPO · CountProb (LLaMA3-8B)win rateNo clear effect
increase KL penalty beta to large values (ablation)LDL-GRPO · CountProb (LLaMA3-8B)win rateHurt
add the two-stage SFT-to-LDL-GRPO switchLDL-GRPO · CountProb, A-OKVQA-RGtraining lossImproved

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