AI Paper Atlas

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

Sparse but Critical: A Token-Level Analysis of Distributional Shifts in RLVR Fine-Tuning of LLMs

Haoming Meng, Kexin Huang, Shaohang Wei, Chiyu Ma, Shuo Yang, Xue Wang, Guoyin Wang, Bolin Ding, Jingren Zhou

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 19 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
inject RL-sampled tokens at high-divergence positions into base generations (forward cross-sampling) (ablation)Qwen2.5-32B base with SimpleRL · AIME 2024Mean@32 accuracy (%): 8.23 -> >25 with 3.86% tokensImproved
inject RL-sampled tokens into base generations (ablation)Qwen2.5-32B base with SimpleRL · AIME 2025Mean@32 accuracy (%): 5.3 -> >14 with 1.53% tokensImproved
inject DAPO tokens into base generations (ablation)Qwen2.5-32B base with DAPO · AIME 2024Mean@32 accuracy (%): 8.23 -> >44 with 7.8% tokensImproved
revert RL tokens to base samples at high-divergence positions (reverse cross-sampling) (ablation)Qwen2.5-32B SimpleRL · AIME 2024Mean@32 accuracy (%): 25.52 -> <8.3 with 5% tokensHurt
revert RL tokens to base samples (ablation)Qwen2.5-32B DAPO · AIME 2025Mean@32 accuracy (%): 32 -> <4.5 with 9.89% tokensHurt
use clip-higher (0.28) instead of 0.2 upper clip in DAPO (ablation)Qwen2.5-Math-7B DAPO · token JS divergenceshare near-zero divergenceMixed
use SFT instead of RLVR (DAPO)Qwen2.5-32B · token JS divergencehigh-divergence set size: sigma 0.00143 (RL) vs 0.00347 (SFT)Hurt
weight advantages up for high-KL tokens (sigmoid weighting) (ablation)DAPO on Qwen2.5-Math-7B · AIME24/AIME25/AMC avgaccuracy (%): 42.48 -> 45.05Improved
weight advantages up for low-KL tokens (ablation)DAPO on Qwen2.5-Math-7B · AIME24/AIME25/AMC avgaccuracy (%): 42.48 -> 44.92Improved
weight advantages up for high-KL tokens (ablation)DAPO on Qwen2.5-Math-7B · AMCaccuracy (%): 75.08 -> 78.40Improved

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