AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Learning What Reinforcement Learning Can't: Interleaved Online Fine-Tuning for Hardest Questions

Lu Ma, Hao Liang, Meiyi Qiang, Lexiang Tang, Xiaochen Ma, Zhen Hao Wong, Junbo Niu, Chengyu Shen, Runming He, Yanhao Li, Wentao Zhang, Bin CUI

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 96 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
interleave online SFT on hardest questions into RLGRPO RL on Qwen2.5-Math-7B · overall six benchmarksaccuracy (%): 46.9 -> 52.6Improved
interleave online SFT on hardest questions vs SFTQwen2.5-Math-7B · overall six benchmarksaccuracy (%): 49.0 -> 52.6Improved
use ReLIFT instead of LUFFYQwen2.5-Math-7B · overallaccuracy (%): 50.9 -> 52.6Improved
use ReLIFT instead of SFT then RLQwen2.5-Math-7B · overallaccuracy (%): 50.8 -> 52.6Improved
add ReLIFT on Qwen2.5-Math-1.5BRL · overallaccuracy (%): 34.2 -> 36.5Improved
add ReLIFT on Qwen2.5-7BRL · overallaccuracy (%): 43.1 -> 45.0Improved
add ReLIFT on Llama3.1-8BRL · overallaccuracy (%): 14.6 -> 17.3Improved
add adaptive scheduling vs fine-tuning after every RL step (ablation)ReLIFT · overallaccuracy (%): 23.8 -> 52.6Improved
add adaptive scheduling vs uniform intervals (ablation)ReLIFT · overallaccuracy (%): 49.1 -> 52.6Improved
select hardest questions instead of random non-hardest (ablation)ReLIFT · overallaccuracy (%): 45.5 -> 52.6Improved
add entropy loss with alpha=1e-4 to fine-tuning (ablation)ReLIFT · Qwen2.5-Math-7Baccuracy: 52.6Improved
increase buffer size from 32 to 64 at Q=0 (ablation)ReLIFT · 5 math benchmarksaverage accuracy: 0.411 -> 0.490Improved

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