Saved in:
Bibliographic Details
Main Authors: Li, Xirui, Li, Ming, Zhou, Tianyi
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.12395
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915795988119552
author Li, Xirui
Li, Ming
Zhou, Tianyi
author_facet Li, Xirui
Li, Ming
Zhou, Tianyi
contents Reinforcement learning (RL) with verifiable rewards has become a standard post-training stage for boosting visual reasoning in vision-language models, yet it remains unclear what capabilities RL actually improves compared with supervised fine-tuning as cold-start initialization (IN). End-to-end benchmark gains conflate multiple factors, making it difficult to attribute improvements to specific skills. To bridge the gap, we propose a Frankenstein-style analysis framework including: (i) functional localization via causal probing; (ii) update characterization via parameter comparison; and (iii) transferability test via model merging. Instead, RL induces a consistent inference-time shift primarily in mid-to-late layers, and these mid-to-late refinements are both transferable (via merging) and necessary (via freezing) for RL gains. Overall, our results suggest that RL's reliable contribution in visual reasoning is not a uniform enhancement of visual perception, but a systematic refinement of mid-to-late transformer computation that improves vision-to-reasoning alignment and reasoning performance, highlighting the limitations of benchmark-only evaluation for understanding multimodal reasoning improvements.
format Preprint
id arxiv_https___arxiv_org_abs_2602_12395
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle What does RL improve for Visual Reasoning? A Frankenstein-Style Analysis
Li, Xirui
Li, Ming
Zhou, Tianyi
Computer Vision and Pattern Recognition
Artificial Intelligence
Reinforcement learning (RL) with verifiable rewards has become a standard post-training stage for boosting visual reasoning in vision-language models, yet it remains unclear what capabilities RL actually improves compared with supervised fine-tuning as cold-start initialization (IN). End-to-end benchmark gains conflate multiple factors, making it difficult to attribute improvements to specific skills. To bridge the gap, we propose a Frankenstein-style analysis framework including: (i) functional localization via causal probing; (ii) update characterization via parameter comparison; and (iii) transferability test via model merging. Instead, RL induces a consistent inference-time shift primarily in mid-to-late layers, and these mid-to-late refinements are both transferable (via merging) and necessary (via freezing) for RL gains. Overall, our results suggest that RL's reliable contribution in visual reasoning is not a uniform enhancement of visual perception, but a systematic refinement of mid-to-late transformer computation that improves vision-to-reasoning alignment and reasoning performance, highlighting the limitations of benchmark-only evaluation for understanding multimodal reasoning improvements.
title What does RL improve for Visual Reasoning? A Frankenstein-Style Analysis
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2602.12395