SFT or RL? An Early Investigation into Training R1-Like Reasoning Large
  Vision-Language Models

Hardy Chen; Haoqin Tu; Fali Wang; Hui Liu; Xianfeng Tang; Xinya Du; Yuyin Zhou; Cihang Xie

返回文献探索

Paper · arXiv 2504.11468

SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models

Hardy Chen, Haoqin Tu, Fali Wang, Hui Liu, Xianfeng Tang, Xinya Du, Yuyin Zhou, Cihang Xie

30 upvotesApril 10, 2025arXiv 预印本

AI 摘要

Supervised fine-tuning can hinder reinforcement learning in Large Vision-Language Models by inducing imitative reasoning, whereas a novel reinforcement learning approach improves genuine reasoning capabilities and outperforms existing models.

supervised fine-tuningreinforcement learningLarge Vision-Language Modelspseudo reasoning pathsVLAA-Thinkingreasoning distillationanswer rewriteverificationGroup Relative Policy Optimizationmixed reward moduleperception signalscognition signalsVLAA-ThinkerQwen2.5VL 3BOpen LMM Reasoning Leaderboard

Abstract

This work revisits the dominant supervised fine-tuning (SFT) then reinforcement learning (RL) paradigm for training Large Vision-Language Models (LVLMs), and reveals a key finding: SFT can significantly undermine subsequent RL by inducing ``pseudo reasoning paths'' imitated from expert models. While these paths may resemble the native reasoning paths of RL models, they often involve prolonged, hesitant, less informative steps, and incorrect reasoning. To systematically study this effect, we introduce VLAA-Thinking, a new multimodal dataset designed to support reasoning in LVLMs. Constructed via a six-step pipeline involving captioning, reasoning distillation, answer rewrite and verification, VLAA-Thinking comprises high-quality, step-by-step visual reasoning traces for SFT, along with a more challenging RL split from the same data source. Using this dataset, we conduct extensive experiments comparing SFT, RL and their combinations. Results show that while SFT helps models learn reasoning formats, it often locks aligned models into imitative, rigid reasoning modes that impede further learning. In contrast, building on the Group Relative Policy Optimization (GRPO) with a novel mixed reward module integrating both perception and cognition signals, our RL approach fosters more genuine, adaptive reasoning behavior. Notably, our model VLAA-Thinker, based on Qwen2.5VL 3B, achieves top-1 performance on Open LMM Reasoning Leaderboard (https://huggingface.co/spaces/opencompass/Open_LMM_Reasoning_Leaderboard) among 4B scale LVLMs, surpassing the previous state-of-the-art by 1.8%. We hope our findings provide valuable insights in developing reasoning-capable LVLMs and can inform future research in this area.