The Synergy Dilemma of Long-CoT SFT and RL: Investigating Post-Training Techniques for Reasoning VLMs

TL;DR

This paper systematically investigates the effects and interactions of long-CoT supervised fine-tuning and reinforcement learning in vision-language models, revealing trade-offs and the need for more adaptive combined training methods.

Contribution

It provides a comprehensive analysis of how long-CoT SFT and RL individually and jointly affect reasoning in VLMs, highlighting their limitations and the complexity of combining them effectively.

Findings

SFT improves performance on difficult questions but increases verbosity and harms simple question accuracy.

RL enhances generalization and maintains brevity, with consistent improvements across question difficulties.

Combining SFT and RL through various strategies does not yield additive benefits and introduces trade-offs.

Abstract

Large vision-language models (VLMs) increasingly adopt post-training techniques such as long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL) to elicit sophisticated reasoning. While these methods exhibit synergy in language-only models, their joint effectiveness in VLMs remains uncertain. We present a systematic investigation into the distinct roles and interplay of long-CoT SFT and RL across multiple multimodal reasoning benchmarks. We find that SFT improves performance on difficult questions by in-depth, structured reasoning, but introduces verbosity and degrades performance on simpler ones. In contrast, RL promotes generalization and brevity, yielding consistent improvements across all difficulty levels, though the improvements on the hardest questions are less prominent compared to SFT. Surprisingly, combining them through two-staged, interleaved, or progressive training strategies, as well as data mixing and model merging, all fails to produce additive benefits, instead leading to trade-offs in accuracy, reasoning style, and response length. This ``synergy dilemma'' highlights the need for more seamless and adaptive approaches to unlock the full potential of combined post-training techniques for reasoning VLMs.