Research map / Vision-language and multimodal / Multimodal LLMs

OpenMMReasoner: Pushing the Frontiers in Multimodal Reasoning with an Open and General Recipe

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces OpenMMReasoner, a fully open two-stage SFT+RL recipe (874K cold-start SFT samples, 74K RL samples) that trains Qwen2.5-VL-7B into a multimodal reasoning model, reporting an 11.6% average gain over the base across nine benchmarks.

Problem

Multimodal reasoning research lacks transparent, reproducible data curation and training pipelines combining SFT and RL, making it hard to know how reasoning-capable LMMs are built.

Approach

A fully released pipeline that scales answer diversity (repeated teacher sampling with verification), mixes cross-domain math data for SFT, and selects GSPO with a large rollout count and no difficulty filtering for RL, with ablations justifying each choice.

Key results

• OMR-7B reaches 79.5 on MathVista testmini versus 69.2 for Qwen2.5-VL-7B and 72.1 for OVR-7B.
• Scaling teacher answer samples from x1 to x8 raises the SFT average on visual reasoning benchmarks from 50.5 to 55.2.
• Text-only reasoning transfers: AIME24 rises from 6.7 (baseline) to 27.1 after RL, with AIME25 at 22.1.

Surprising findings

• Filtering SFT data by length or difficulty reduced performance, since it lowered answer diversity.
• Multimodal-only RL also improved text-only math benchmarks (AIME24/25, AMC23), indicating cross-modal transfer.
• Curriculum sampling did not beat the original mixed sampling in RL.

Limitations

• Focuses on a single model family (Qwen2.5-VL-Instruct) and mainly image-domain evaluation.
• Does not extend to video or audio modalities.
• Upper bound of performance under further scaling not identified.

Experimental setup

Datasets: LLaVA-CoT OpenVLThinker We-Math 2.0 MMR1 MiroMind-M1 MMEureka ViRL TQA We-Math PuzzleVQA AlgoPuzzleVQA ThinkLiteVL

Benchmarks: MathVista MathVision MathVerse DynaMath WeMath LogicVista MMMU MMMU-Pro CharXiv AIME24 AIME25 GPQA Diamond

Metrics: accuracy average accuracy

Baselines: Qwen2.5-VL-7B-Instruct OpenVisionReasoner (OVR-7B) M2-Reasoning MMR1 VL-Rethinker-7B MM-Eureka-Qwen-7B OpenVLThinker-7B ThinkLite-7B-VL LLaVA-OneVision-1.5-8B InternVL3-8B

Builds on: GRPO DAPO GSPO Qwen2.5-VL

Closest papers by meaning

• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.92)
• Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning (NeurIPS 2025 · similarity 0.90)
• RPTS: Tree-Structured Reasoning Process Scoring for Faithful Multimodal Evaluation (AAAI 2026 · similarity 0.90)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.90)
• OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025 · similarity 0.90)
• MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning (ICLR 2026 · similarity 0.89)