ProcessThinker boosts multimodal reasoning without explicit reward model

A team of researchers from multiple institutions has introduced ProcessThinker, a novel post-training pipeline designed to improve reasoning in multimodal large language models (LLMs) without the cost of training an explicit process reward model (PRM). As visual question answering increasingly demands multi-step logical reasoning, existing reinforcement learning methods like Group Relative Policy Optimization (GRPO) often rely on sparse outcome-only rewards, making it hard to pinpoint whether an incorrect answer stems from a small late mistake or a fundamentally flawed reasoning trajectory.

ProcessThinker addresses this by first rewriting reasoning traces into a step-tagged format for cold-start supervised fine-tuning, then applying GRPO with a rollout-based process reward. For each intermediate step, the method samples multiple continuations and uses the empirical success rate—verified against the final answer—as the step reward. This dense credit assignment encourages reasoning steps that reliably support a correct conclusion, reducing self-contradictory progress. Across four challenging video benchmarks (Video-MMMU, MMVU, VideoMathQA, LongVideoBench), the pipeline consistently improves over the baseline Qwen3-VL-8B-Instruct, demonstrating practical gains in multimodal logical reasoning.