Source-linked AI summary
Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought
Yi Peng, Peiyu Wang, Xiaokun Wang, Yichen Wei, Jiangbo Pei, Weijie Qiu, Ai Jian, Yunzhuo Hao, Jiachun Pan, Tianyidan Xie, Li Ge, Rongxian Zhuang, Xuchen Song, Yang Liu, Yahui Zhou
TL;DR
Skywork R1V addresses the difficulty of extending strong textual reasoning to complex visual tasks and the scarcity of suitable multimodal reasoning data. It uses an MLP-based transfer method, hybrid SFT and GRPO optimization, and adaptive-length CoT distillation. The 38B-parameter model achieves competitive visual performance while retaining strong textual reasoning and publicly releasing its weights.
Problem
Existing vision-language models struggle with deeply logical multimodal tasks, while available datasets provide limited complex reasoning-oriented training data.
Method
Skywork R1V transfers R1-series reasoning through an MLP adapter, combines iterative SFT with GRPO, and uses adaptive-length CoT distillation for reasoning-data generation.
Results
Skywork R1V shows strong cross-task performance, scoring 69.0 on MMMU, 67.5 on MathVista, 72.0 on AIME 2024, and 94.0 on MATH-500.
Takeaways & Limitations
The results indicate that the approach transfers advanced textual reasoning to vision while enabling a 38B-scale model to perform competitively with larger models.
Abstract
from arXiv · showhide
We introduce Skywork R1V, a multimodal reasoning model extending the an R1-series Large language models (LLM) to visual modalities via an efficient multimodal transfer method. Leveraging a lightweight visual projector, Skywork R1V facilitates seamless multimodal adaptation without necessitating retraining of either the foundational language model or the vision encoder. To strengthen visual-text alignment, we propose a hybrid optimization strategy that combines Iterative Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO), significantly enhancing cross-modal integration efficiency. Additionally, we introduce an adaptive-length Chain-of-Thought distillation approach for reasoning data generation. This approach dynamically optimizes reasoning chain lengths, thereby enhancing inference efficiency and preventing excessive reasoning overthinking. Empirical evaluations demonstrate that Skywork R1V, with only 38B parameters, delivers competitive performance, achieving a score of 69.0 on the MMMU benchmark and 67.5 on MathVista. Meanwhile, it maintains robust textual reasoning performance, evidenced by impressive scores of 72.0 on AIME and 94.0 on MATH500. The Skywork R1V model weights have been publicly released to promote openness and reproducibility.
1 Introduction
Skywork R1V extends R1-series reasoning into visual tasks through efficient multimodal transfer, hybrid optimization, and adaptive-length reasoning distillation. It targets weaknesses in existing vision-language reasoning while keeping model components open-sourced.
- Existing vision-language models often underperform on deeply logical multimodal tasks such as geometric proofs and scientific problem-solving.
- Current multimodal datasets contain limited complex reasoning data, so training on predominantly non-reasoning content can weaken multimodal reasoning capabilities.
- Skywork R1V transfers R1-series textual reasoning to visual tasks with a lightweight MLP projector without retraining the base language model or visual encoder.
- Its hybrid optimization framework combines Iterative Supervised Fine-Tuning with Group Relative Policy Optimization to align visual and textual representations.
- Adaptive-length Chain-of-Thought distillation dynamically adjusts reasoning-chain lengths to mitigate excessive computational deliberation and improve inference efficiency.
- All Skywork R1V components and weights are fully open-sourced to support research and innovation in multimodal reasoning.
2 Methodology
The methodology stages multimodal transfer, iterative hybrid optimization, and adaptive-length reasoning distillation to align vision with a reasoning-capable language model efficiently. The design preserves the language model and vision encoder while refining an MLP adapter and controlling generated reasoning length.
- Efficient Multimodal Transfer: The efficient transfer method aligns a vision backbone with a reasoning-capable language model through an MLP structure, reducing the need for extensive multimodal reasoning data.
- Efficient Multimodal Transfer: The pretrained MLP adapter is reassembled between the vision encoder and reasoning-capable language model, while both components remain fixed during modality alignment.
- Hybrid Optimization Framework: Hybrid optimization combines iterative SFT using reward-selected and previously misclassified samples with GRPO reinforcement learning for generalization.
- Hybrid Optimization Framework: The customized iterative-training dataset combines high-quality reward-model samples with error-focused samples from the preceding model iteration.
- Adaptive-Length Chain-of-Thought Distillation: AL-CoTD dynamically optimizes reasoning-chain length during reasoning-data generation to mitigate excessive reasoning or overthinking at inference.
3 Experiments
Skywork R1V is evaluated across textual reasoning, visual multimodal benchmarks, larger-scale model comparisons, and staged training analyses. Results show strong performance across both reasoning and visual tasks, with iterative SFT and GRPO improving MMMU performance.
- Main Results: 94.0 on MATH-500 and 72.0 on AIME 2024 demonstrate strong textual reasoning performance.Skywork R1V surpasses QwQ-32B-Preview’s 90.6 on MATH-500.
- Main Results: 69.0 Pass@1 on MMMU (Val) surpasses several similar-scale models, while 67.5 on MathVista trails InternVL2.5-38B’s 71.9.On MMMU, Skywork R1V exceeds VILA-40B, InternVL2-40B, InternVL2.5-38B, and QwenVL2-72B.
- Main Results: 67.5 on MathVista surpasses Claude 3.5 Sonnet’s 65.3, while 69.0 on MMMU outperforms Claude 3.5 Sonnet’s 66.4 and matches GPT4-o’s 69.1.These comparisons concern larger-scale open-source and closed-source models.
- Reasoning Capability: Skywork R1V uses structured mathematical reasoning and verification to solve a pentagon angle problem, obtaining x = 102 and H = 97°.The model constructs and solves a linear equation, then checks the angle sum by substitution.
- Reasoning Capability: On a chart task, the model computes yearly female–male life-expectancy differences, cross-checks them against visual trends, and identifies 2019 as the peak disparity year.The analysis covers U.S. life expectancy data from 2009–2019 and includes recalculation for error checking.
- Training Analysis: MMMU performance rises from 60.2 after model re-assembly to 65.6 after the fifth iterative-SFT stage and 69.0 after GRPO.The staged results indicate successive gains during iterative SFT followed by a further increase after reinforcement learning.