PaperlayerView exact revision on arXiv

Source-linked AI summary

MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods

Honglin Lin, Zheng Liu, Yun Zhu, Chonghan Qin, Juekai Lin, Xiaoran Shang, Conghui He, Wentao Zhang, Lijun Wu

arXiv:2601.21821v1cs.CV

TL;DR

Open-source VLMs lack multimodal reasoning data with sufficient coverage, consistency, and long-form supervision. MMFineReason constructs a 1.8M-sample, 5.1B-token dataset through distilled reasoning annotation and difficulty-aware selection, then fine-tunes compact models that outperform larger baselines while a 7% subset matches full-data performance.

  • Problem

    Open multimodal datasets lack sufficient STEM and visual-puzzle coverage and consistent, interpretable long-form Chain-of-Thought supervision.

  • Method

    MMFineReason aggregates and standardizes data, distills visually grounded reasoning traces from Qwen3-VL-235B-A22B-Thinking, and applies quality and difficulty-aware selection before fine-tuning Qwen3-VL-Instruct.

  • Results

    MMFineReason-4B surpasses Qwen3-VL-8B-Thinking, MMFineReason-8B outperforms Qwen3-VL-30B-A3B-Thinking, and a 7% subset achieves performance comparable to the full dataset.

  • Takeaways & Limitations

    Reasoning-oriented data composition improves both specialized reasoning and general capabilities, enabling smaller open models to compete with larger reasoning-oriented systems.

  • Takeaways & Limitations

    RL improves DynaMath across scales but causes slight MathVision drops for 4B and 8B models, leaving more effective RL data strategies for future work.

Abstract

from arXiv · show

Recent advances in Vision Language Models (VLMs) have driven significant progress in visual reasoning. However, open-source VLMs still lag behind proprietary systems, largely due to the lack of high-quality reasoning data. Existing datasets offer limited coverage of challenging domains such as STEM diagrams and visual puzzles, and lack consistent, long-form Chain-of-Thought (CoT) annotations essential for eliciting strong reasoning capabilities. To bridge this gap, we introduce MMFineReason, a large-scale multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring high-quality reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking. The dataset is established via a systematic three-stage pipeline: (1) large-scale data collection and standardization, (2) CoT rationale generation, and (3) comprehensive selection based on reasoning quality and difficulty awareness. The resulting dataset spans STEM problems, visual puzzles, games, and complex diagrams, with each sample annotated with visually grounded reasoning traces. We fine-tune Qwen3-VL-Instruct on MMFineReason to develop MMFineReason-2B/4B/8B versions. Our models establish new state-of-the-art results for their size class. Notably, MMFineReason-4B succesfully surpasses Qwen3-VL-8B-Thinking, and MMFineReason-8B even outperforms Qwen3-VL-30B-A3B-Thinking while approaching Qwen3-VL-32B-Thinking, demonstrating remarkable parameter efficiency. Crucially, we uncover a "less is more" phenomenon via our difficulty-aware filtering strategy: a subset of just 7\% (123K samples) achieves performance comparable to the full dataset. Notably, we reveal a synergistic effect where reasoning-oriented data composition simultaneously boosts general capabilities.

1 Introduction

MMFineReason addresses multimodal reasoning data gaps through a large-scale, distilled, and difficulty-aware dataset pipeline. Fine-tuned models achieve strong parameter-efficient performance, while reasoning-oriented composition and small filtered subsets retain broad gains.

  • Motivation: Open multimodal datasets remain limited by scarce high-quality STEM and visual-puzzle samples and inconsistent long-form Chain-of-Thought supervision.
  • Dataset: MMFineReason contains over 1.8M samples and 5.1B solution tokens with reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
  • Method: The dataset pipeline aggregates and standardizes multimodal data, distills visually grounded reasoning traces, and selects examples using quality verification and difficulty-based filtering.
  • Results: MMFineReason-4B surpasses Qwen3-VL-8B-Thinking, while MMFineReason-8B outperforms Qwen3-VL-30B-A3B-Thinking and approaches Qwen3-VL-32B-Thinking.
  • Results: Reasoning-oriented data spanning STEM, puzzles, games, and diagrams improves both specialized reasoning and general model performance.
  • Efficiency: A 7% filtered subset achieves performance comparable to training on the full dataset, supporting efficient fine-tuning through difficulty-aware selection.

Overall Findings

MMFineReason reports strong scaling efficiency but also identifies substantial imbalance in training-data difficulty and domain composition.

  • 5.1B tokens suffice to surpass Qwen3-VL-30B-A3B-thinking and achieve performance comparable to GPT-5-mini-high with a strong pretrained backbone.
  • 67% of current training data is overly simple, while puzzle-style problems are significantly harder.

2 Related Work

Prior multimodal reasoning datasets provide important resources but remain limited in coverage, quality, transparency, and reproducibility. MMFineReason responds with broader collection and a reproducible curation and training pipeline.

  • Multimodal Reasoning Datasets: Multimodal reasoning covers mathematical problem solving, visual logical reasoning, and chart understanding, but open-source datasets remain smaller than the substantial private datasets used by proprietary models.
  • Multimodal Reasoning Datasets: MathV360K and LLaVA-CoT primarily emphasize mathematical reasoning, limiting coverage, while FineVision aggregates data with relatively coarse and lower-quality sources.
  • Data Construction: Earlier pipelines used strong vision-language models for visual instruction, image description, filtering, and synthetic-data generation.
  • Data Construction: MMFineReason introduces a high-quality, fully reproducible curation and training pipeline intended to help open-source multimodal models narrow the gap with closed-source systems.

3 MMFineReason Pipeline

MMFineReason is constructed through collection, cleaning, standardization, teacher-based annotation, and multi-stage selection. The resulting dataset combines visually grounded reasoning traces with automated quality and difficulty signals for efficient training.

  • Data Collection: The pipeline aggregates multimodal sources, manually filters for STEM and reasoning relevance, and expands coverage of mathematical, scientific, game, and puzzle tasks.Sources include FineVision, BMMR, Euclid30K, Zebra-CoT-Physics, and GameQA-140K.
  • Data Cleaning: Data cleaning standardizes language, removes textual and image artifacts, refines shallow prompts, and excludes tasks outside visual analytical reasoning.Non-English questions are translated, noisy formatting is removed, prompts encourage careful reasoning, and unsuitable coding or drawing tasks are filtered.
  • Data Standardization: The standardized schema preserves source identifiers and original question-answer fields while organizing training inputs, augmented annotations, and verification metrics.Teacher-generated captions and reasoning responses are paired with pass rate, consistency, and consistency-analysis metrics.
  • Reasoning Distillation: Qwen3-VL-235B-A22B-Thinking generates long-form reasoning traces through a four-phase framework, while Qwen3-VL-235B-A22B-Instruct generates dense image captions.Outputs use <think> and <answer> blocks to support downstream parsing and automatic verification.
  • Dataset Construction: The initial MMFineReason-Full corpus contains 2.3M samples and 8.8B solution tokens before quality filtering produces MMFineReason-1.8M with 5.1B solution tokens.A further 40K instances are reserved for reinforcement-learning training, with the remaining data used for supervised fine-tuning.
  • Data Selection: Quality selection validates output structure and length, removes repetitive traces, verifies final answers against ground truth, and filters difficulty using Qwen3-VL-4B-Thinking pass rates.Correctness verification eliminates roughly 20% of instances containing potential hallucinations or incorrect reasoning traces; difficulty filtering discards examples solved in at least one of four attempts.

4 MMFineReason Dataset

MMFineReason is a reasoning-dominant multimodal dataset emphasizing mathematics, science, puzzles, games, and diagrammatic content. Its annotations are longer and denser than prior datasets, with difficulty and visual-semantic analyses exposing both its specialization and scope boundary.

  • Dataset Composition: 1,770,926 samples are distributed across Mathematics (79.4%), Science (13.8%), Puzzle/Game (4.6%), and General/OCR (2.2%).The composition is strategically weighted toward symbolic and logic-heavy reasoning tasks.
  • Dataset Composition: General/OCR contributes 2.2% of the dataset and serves as a regularization set preserving broad visual and OCR capabilities without diluting reasoning-focused supervision.The set contains 38.7K general-purpose samples from LLaVA-CoT.
  • Difficulty Distribution Analysis: ScienceQA, AI2D, and TQA have relatively high pass rates, whereas puzzle and game datasets are harder because they require an unbroken chain of deduction.The cited science-oriented datasets use clean diagrams and mostly multiple-choice formats, while logic and puzzle tasks are sensitive to single-step failures.
  • Response Analysis: 609-token captions average across all image-question pairs, exceeding HoneyBee’s 299-token captions, while MMFineReason provides 100% caption coverage.The captions supply fine-grained visual details for every reasoning chain rather than only a subset of samples.
  • Visual Semantic Analysis: STEM and diagrammatic images account for 75.3% of the corpus, while the natural-image subset spans urban scenes, documents, and astronomical visualizations.The paper prioritizes fine-grained mathematical reasoning while retaining a complementary natural subset for broader generalization assessment.
  • Response Analysis: MMFineReason averages 2,910 CoT tokens, compared with 1,063 for HoneyBee and 675 for OpenMMReasoner.Its median is 2,038 tokens, versus 972 for HoneyBee and 180 for OpenMMReasoner.
  • Response Analysis: Average reasoning lengths vary by domain: Puzzle & Game 4,810 tokens, Mathematics 2,950, Science 2,305, and General/OCR 1,262.Puzzle and game responses emphasize visual-spatial verification, mathematics uses symbolic derivations, science combines perception with domain knowledge, and General/OCR favors direct grounding.

5 Experiments

MMFineReason models achieve strong multimodal reasoning performance with notable data and parameter efficiency, while ablations show distinct roles for supervised fine-tuning, reinforcement learning, and data composition.

  • Main Results: MMFineReason models establish new state-of-the-art results for their size class, with MFR-4B surpassing Qwen3-VL-8B-Thinking and MFR-8B outperforming Qwen3-VL-30B-A3B-Thinking.MFR-8B approaches Qwen3-VL-32B-Thinking while using fewer parameters.
  • Mathematical & Logical Reasoning: 83.4% on DynaMath places MFR-8B above Qwen3-VL-32B-Thinking at 82.0% and Qwen3-VL-30B-A3B-Thinking at 76.7%.On MathVerse, MFR-8B reaches 81.5%, approaching Qwen3-VL-32B-Thinking at 82.6% and surpassing Qwen3-VL-30B-A3B-Thinking at 79.6%.
  • Generalization Across Domains: 75.6% on RWQA and 89.9% on CharXivdesc. show that MFR-8B generalizes competitively to general understanding and chart reasoning.The training data contains minimal chart or real-world samples, yet the models maintain strong performance in these domains.
  • Data Efficiency: 67.1% on MathVision lets MFR-8B outperform HoneyBee-8B at 37.4% and OMR-7B at 36.6%, while 81.5% on MathVerse exceeds MMR1-8B at 67.3% and HoneyBee-8B at 60.9%.These comparisons are presented as evidence that MFR reasoning-chain quality outperforms scale-focused strategies.
  • Training Stages: SFT raises the 8B model from 53.90% to 67.56% on MathVision and from 55.30% to 68.68% on LogicVista.Smaller models also gain after SFT, including +3.5% on MathVerse and +2.8% on LogicVista for the 2B model.
  • Training Stages: RL improves generalization, including AI2D from 78.47% to 82.51% and CharXivreas. from 38.96% to 45.38% for the 2B model, but causes slight MathVision drops for 4B and 8B.The authors attribute remaining RL gains on mathematics to the need for more diverse or challenging data.
  • Scaling and Data Efficiency: 75.7 is MMFineReason-1.8M’s peak score, exceeding MMR1-1.6M at 67.4 and HoneyBee-2.5M at 65.1 despite using fewer samples.The MMFineReason-123K subset reaches 73.3 using approximately 5% of comparable benchmark data volume.
  • Scaling and Data Efficiency: 73.9 from the 4B model surpasses Qwen3-VL-8B-Thinking at 72.5, while high-density datasets such as WeMath2.0-SFT reach 70.98% with only 814 samples.The analysis also reports weaker transfer from several puzzle/game datasets and stronger performance from formalized, diverse reasoning data.

6 Conclusion

MMFineReason frames multimodal reasoning as a data-centric problem and presents structured data design as a way to induce strong reasoning without excessive model scaling or proprietary supervision.

  • Conclusion: MMFineReason treats multimodal reasoning primarily as a data-centric problem rather than a purely model-centric one.The framework covers reasoning supervision, difficulty-aware structuring, and data composition.
  • Conclusion: Structured data design can systematically induce strong multimodal reasoning without relying on excessive model scaling or proprietary supervision.

A Data Curation Details

The curation process prioritizes pre-filtered datasets and excludes data judged unsuitable for reasoning-focused multimodal training.

  • Dataset Selection: Pre-filtered versions are prioritized when available, including FineVision-filtered VisualWebInstruct and OpenMMReasoner-filtered MMR1 subsets.
  • Excluded Categories: Multi-image samples are excluded from the collection.
  • Excluded Categories: Overly simple task data, including CLEVR-family and geo170k data, is excluded because it offers limited reasoning complexity.
  • Excluded Categories: Highly specialized imaging domains such as PathVQA are excluded from the collection.
  • Excluded Categories: Multilingual datasets featuring minor languages within images, such as EXAMS-V, are excluded.

A.3 Filteration Details

MMFineReason applies structural cleaning, semantic consistency checks, and difficulty analysis to retain usable and reliable multimodal reasoning samples.

  • Filtering Pipeline: The processing pipeline combines basic structural cleaning with advanced quality assessment.The reported statistics are summarized in Tables 6 and 7.
  • Basic Data Cleaning: Length filtering removes overlong or anomalously short samples, while template validation discards parsing errors and malformed templates.Most raw datasets retain more than 95% of samples after structural filtering.
  • Consistency and Difficulty Analysis: Pass Rate and Consistency Rate are used together to assess semantic quality, with higher Pass Rate indicating successful model processing.The analysis reports a positive correlation between the two measures.
  • Consistency and Difficulty Analysis: 96.18% consistency for ScienceQA and 97.30% for MMR1 accompany mean Pass Rates of approximately 0.8–0.9.These datasets are characterized as well-posed with clear ground truths.
  • Consistency and Difficulty Analysis: Raven and VisualSphinx have mean Pass Rates below 0.25 and consistency below 40%, indicating challenging or potentially ambiguous samples.The authors note that such tasks require more robust filtering strategies.
  • Final Retention: Approximately 1.81 million consistent samples are retained from the processed pool for stable model training.

B.1 Training Details

Training uses separate supervised fine-tuning and reinforcement-learning stages with specified efficiency, resolution, stability, rollout, and length settings.

  • SFT: SFT uses a cosine scheduler with peak learning rate 1e-5, sequence packing length 32,768, Liger Kernel support, and a 768 × 768 pixel limit.
  • RL: RL follows SFT and uses a constant learning rate of 1e-6 with 16 rollouts per prompt.
  • RL: The RL stage applies a KL-divergence penalty with ϵlow = 3e-4 and ϵhigh = 4e-4.
  • RL: RL input and output maximum lengths are 8,192 and 16,384 tokens, respectively.

B.2 Evaluation Details

Evaluation combines LLM-based correctness verification with standardized multimodal prompting and question-cleaning procedures.

  • Evaluation: Evaluation uses VLMEvalKit with compass-verifier, replacing exact string matching with LLM-as-a-Judge correctness assessment.
  • Evaluation: Rollouts use temperature 0.0, top-p 1.0, top-k −1, repetition penalty 1.05, and maximum response length 32768.
  • Question Cleaning: Question cleaning translates non-English questions and rewrites instructions that discourage reasoning into prompts encouraging a clear reasoning process.
  • Multimodal Annotation: Multimodal annotation requires category-first analysis, separation of visible and inferred content, atomic component descriptions, and holistic use of image and question text.
  • Error Filtering: The pipeline removes irrelevant image-caption and URL content when the visual image already supplies the needed diagram information.
Loading 2601.21821v1…