SophiaVL-R1: Reinforcing MLLMs Reasoning with Thinking Reward

TL;DR

SophiaVL-R1 enhances multimodal large language models by incorporating a thinking reward mechanism with trust-based weighting and annealing strategies, leading to improved reasoning and generalization on multiple benchmarks.

Contribution

It introduces a novel thinking reward model with trust weighting and annealing to improve reasoning in multimodal LLMs, outperforming larger models.

Findings

Outperforms existing reasoning MLLMs on benchmarks like MathVisita and MMMU.

SophiaVL-R1-7B surpasses larger models like LLaVA-OneVision-72B.

Effective mitigation of reward hacking through Trust-GRPO.

Abstract

Recent advances have shown success in eliciting strong reasoning abilities in multimodal large language models (MLLMs) through rule-based reinforcement learning (RL) with outcome rewards. However, this paradigm typically lacks supervision over the thinking process leading to the final outcome. As a result, the model may learn sub-optimal reasoning strategies, which can hinder its generalization ability. In light of this, we propose SophiaVL-R1, as an attempt to add reward signals for the thinking process in this paradigm. To achieve this, we first train a thinking reward model that evaluates the quality of the entire thinking process. Given that the thinking reward may be unreliable for certain samples due to reward hacking, we propose the Trust-GRPO method, which assigns a trustworthiness weight to the thinking reward during training. This weight is computed based on the thinking reward comparison of responses leading to correct answers versus incorrect answers, helping to mitigate the impact of potentially unreliable thinking rewards. Moreover, we design an annealing training strategy that gradually reduces the thinking reward over time, allowing the model to rely more on the accurate rule-based outcome reward in later training stages. Experiments show that our SophiaVL-R1 surpasses a series of reasoning MLLMs on various benchmarks (e.g., MathVisita, MMMU), demonstrating strong reasoning and generalization capabilities. Notably, our SophiaVL-R1-7B even outperforms LLaVA-OneVision-72B on most benchmarks, despite the latter having 10 times more parameters. All code, models, and datasets are made publicly available at https://github.com/kxfan2002/SophiaVL-R1.