Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start

TL;DR

This paper introduces a two-stage training approach combining supervised fine-tuning and reinforcement learning to significantly improve multimodal reasoning in large language models, achieving state-of-the-art results.

Contribution

The study demonstrates that a cold start with structured reasoning patterns followed by RL refinement enhances multimodal reasoning performance.

Findings

Outperforms SFT-only and RL-only methods on benchmarks

Achieves state-of-the-art results for open-source MLLMs at 3B and 7B scales

7B model shows substantial improvements over base models

Abstract

Recent advancements in large language models (LLMs) have demonstrated impressive chain-of-thought reasoning capabilities, with reinforcement learning (RL) playing a crucial role in this progress. While "aha moment" patterns--where models exhibit self-correction through reflection--are often attributed to emergent properties from RL, we first demonstrate that these patterns exist in multimodal LLMs (MLLMs) prior to RL training but may not necessarily correlate with improved reasoning performance. Building on these insights, we present a comprehensive study on enhancing multimodal reasoning through a two-stage approach: (1) supervised fine-tuning (SFT) as a cold start with structured chain-of-thought reasoning patterns, followed by (2) reinforcement learning via GRPO to further refine these capabilities. Our extensive experiments show that this combined approach consistently outperforms both SFT-only and RL-only methods across challenging multimodal reasoning benchmarks. The resulting models achieve state-of-the-art performance among open-source MLLMs at both 3B and 7B scales, with our 7B model showing substantial improvements over base models (e.g., 66.3 %$\rightarrow$73.4 % on MathVista, 62.9 %$\rightarrow$70.4 % on We-Math) and our 3B model achieving performance competitive with several 7B models. Overall, this work provides practical guidance for building advanced multimodal reasoning models. Our code is available at https://github.com/waltonfuture/RL-with-Cold-Start.