ViSurf: Visual Supervised-and-Reinforcement Fine-Tuning for Large Vision-and-Language Models

Post-training Large Vision-and-Language Models (LVLMs) typically involves Supervised Fine-Tuning (SFT) for knowledge injection or Reinforcement Learning with Verifiable Rewards (RLVR) for performance enhancement. However, SFT often leads to sub-optimal performance, while RLVR remains constrained by the model's internal knowledge base. While a sequential SFT $\rightarrow$ RLVR pipeline can be used, it introduces significant computational overhead and suffers from catastrophic forgetting. To address these limitations, we propose ViSurf (\textbf{Vi}sual \textbf{Su}pervised-and-\textbf{R}einforcement \textbf{F}ine-Tuning), a unified, single-stage paradigm that integrates the strengths of both SFT and RLVR. By analyzing their training objectives, we establish a unified framework that injects ground-truth labels directly into RLVR rollouts, facilitating simultaneous external supervision and internal reinforcement. Furthermore, we introduce three novel reward control strategies to ensure training stability and optimization. Extensive experiments demonstrate that ViSurf consistently outperforms standalone SFT, RLVR, and the traditional two-stage pipeline across diverse benchmarks. In-depth analysis corroborates these findings, validating the derivation and design principles of ViSurf.

To teach model to understand images and text, researchers traditionally use two different training methods: one acts like a strict teacher showing the exact answers (Supervised Fine-Tuning), while the other lets the model learn through trial and error (Reinforcement Learning). Unfortunately, the "teacher" method often makes the model forget what it previously learned, while the "trial and error" method fails if the model lacks a baseline understanding of the task. Combining them sequentially is computationally slow and still leads to forgetfulness. To solve this, we developed ViSurf, a unified training approach that blends both methods into a single, seamless step. ViSurf acts like an adaptive tutor: it lets the model explore and learn from its own attempts, but seamlessly steps in with the exact correct answer whenever the model gets stuck. This unified approach allows the model to learn new, complex visual tasks, like identifying medical anomalies or solving visual math problems, without forgetting its foundational knowledge. Ultimately, ViSurf creates more reliable model that significantly outperform previous standalone training methods.