Research map / Vision-language and multimodal / Multimodal LLMs

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

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

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.

Builds MMFineReason, a 1.8M-sample multimodal reasoning dataset with long CoT distilled from Qwen3-VL-235B-A22B-Thinking, and fine-tunes Qwen3-VL-Instruct (2B/4B/8B) with SFT and RL to reach open-source SOTA.

Problem

Open-source VLMs lag proprietary systems on multimodal reasoning, due to scarce STEM/puzzle reasoning data and inconsistent, short, or missing chain-of-thought annotations.

Approach

Aggregate and clean open multimodal STEM/puzzle datasets, distill long four-phase CoT traces from a strong open VLM teacher, filter by template, length, n-gram repetition and answer correctness, and use pass-rate difficulty filtering to build efficient subsets.

Key results

• MMFineReason-8B averages 75.7 across benchmarks, above Qwen3-VL-30B-A3B-Thinking (74.5) and Gemini-2.5-Flash (75.0), approaching Qwen3-VL-32B-Thinking (77.9).
• MMFineReason-4B averages 73.9, surpassing Qwen3-VL-8B-Thinking (72.5).
• Fine-tuning on the 123K difficulty-filtered subset reaches 73.3 average at 8B, versus 67.4 for MMR1 and 65.1 for HoneyBee trained on full data.

Surprising findings

• Caption augmentation gives marginal or negative gains on STEM benchmarks once long CoT is present.
• Ultra-high resolution (2048^2) does not outperform 768^2 on most reasoning benchmarks, though it helps RealWorldQA.

Limitations

• RL shows variance on math benchmarks, with slight MathVision drops for 4B and 8B.
• Evaluation relies on a single greedy-decoding setup and LLM-based answer verification (CompassVerifier)
• Mixing ratios between STEM and natural image data are left to future work; data is dominated by mathematics (79.4%).

Experimental setup

Datasets: MMFineReason-1.8M FineVision MMR1 BMMR Euclid30K Zebra-CoT-Physics GameQA-140K VisualWebInstruct LLaVA-CoT WaltonColdStart ViRL39K MMK12

Benchmarks: MMMU MathVista MathVision MathVerse DynaMath LogicVista VisuLogic ScienceQA RealWorldQA MMBench-EN MMStar AI2D

Metrics: accuracy

Baselines: Gemini-2.5-Flash GPT5-mini High Qwen3-VL-8B-Thinking Qwen3-VL-30B-A3B-Thinking Qwen3-VL-32B-Thinking MMR1-8B HoneyBee-8B OMR-7B

Builds on: FineVision Qwen3-VL GSPO LLaMA-Factory VeRL

Closest papers by meaning

• Long Grounded Thoughts: Synthesizing Visual Problems and Reasoning Chains at Scale (ICML 2026 · similarity 0.93)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.93)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.93)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.92)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.92)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.92)