AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Why Does RL Generalize Better Than SFT? A Data-Centric Perspective on VLM Post-Training

Aojun Lu, Tao Feng, Hangjie Yuan, Wei Li, Yanan Sun

CVPR 2026 · Reinforcement learning › Visual and representation RL · approach: Robust visual encoders for RL · 3 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
train SFT on hard-difficulty samples only (ablation)Qwen2.5-VL-7B SFT · ImageNet-Raccuracy: 57.32 -> 43.25Hurt
train SFT on hard-difficulty samples only (ablation)Qwen2.5-VL-3B SFT · ImageNet-Raccuracy: 60.51 -> 51.17Hurt
train SFT on medium-difficulty samples only (ablation)Qwen2.5-VL-7B SFT · ImageNet (ID)accuracy: 78.26 -> 85.52Improved
train SFT on easy samples only (ablation)Qwen2.5-VL-7B SFT · ImageNet-Raccuracy: 57.32 -> 59.37Improved
filter to easy plus medium samples (SFT-EM) instead of full dataSFT · OOD average (7B)accuracy: 57.62 -> 62.10Improved
train on medium samples only (SFT-M) instead of full-data SFTSFT · OOD average (7B)accuracy: 57.62 -> 60.41Improved
DC-SFT variants instead of full-data SFTSFT · ImageNet (ID)accuracy: 93.70 -> 91.44Hurt
SFT-M instead of GRPOGRPO · OOD average (7B)accuracy: 59.48 -> 60.41Improved
SFT-EM instead of GRPOGRPO · OOD average (7B)accuracy: 59.48 -> 62.10Improved
add 5% hard data to SFT-EM data (ablation)SFT-EM · ImageNet-Raccuracy: -3.74Hurt
add 5% hard data to SFT-EM data (ablation)SFT-EM · ImageNet-Aaccuracy: -2.51Hurt
DC-SFT instead of GRPO trainingGRPO · training efficiencyspeedup: 4.9x on ImageNet, 3.2x on RefCOCOImproved

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