Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought (2504.05599v2)

Abstract: We introduce Skywork R1V, a multimodal reasoning model extending the an R1-series LLMs (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 LLM 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.

Skywork R1V: Advancements in Multimodal Reasoning

The paper "Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought" details the development of a novel multimodal reasoning model, Skywork R1V, which extends the capabilities of the R1-series LLMs to include visual modalities. This paper introduces several innovative techniques for achieving efficient multimodal transfer and enhanced reasoning abilities, presenting empirical results that demonstrate the competitive performance of the Skywork R1V model.

Core Innovations

Skywork R1V integrates advanced reasoning functionalities into visual modalities through the use of a lightweight visual projector and other enhancements. The key innovations highlighted in the paper are:

1. Efficient Multimodal Transfer: Utilizing a lightweight multilayer perceptron (MLP) as a visual projector allows for the seamless integration of reasoning capabilities from text-based LLMs into visual contexts without retraining the foundational models.
2. Hybrid Optimization Framework: This framework combines Iterative Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO), facilitating the alignment of visual and textual representations and promoting efficient cross-modal reasoning.
3. Adaptive-Length Chain-of-Thought (CoT) Distillation: This approach dynamically adjusts reasoning chain lengths to improve inference efficiency, minimizing excessive computational overthinking.

Methodological Approach

The paper describes a three-step Efficient Multimodal Transfer method that significantly reduces the need for extensive multimodal reasoning datasets. Initially, a vision encoder is aligned with a substitutive LLM using an MLP, followed by transferring and optimizing this pretrained MLP to align with the original reasoning-capable LLM. The alignment process utilizes hybrid optimization techniques and reasoning chains generated dynamically.

Experimental Evaluation

Skywork R1V's performance is empirically validated across multiple benchmarks that test reasoning and visual capabilities:

• Reasoning Tasks: Skywork R1V achieved remarkable results with scores of 94.0 on MATH-500 and 72.0 on AIME 2024, showcasing its strong reasoning abilities.
• Visual Multimodal Tasks: It also performed competitively on the MathVista and MMMU benchmarks, achieving scores of 67.5 and 69.0 respectively, indicating its effective adaptation to multimodal inputs.

Implications and Future Directions

The research highlights significant potential for integrating advanced reasoning capabilities into multimodal AI systems, bridging the gap between text-based LLMs and vision-LLMs. The efficient transfer methods and techniques developed can inspire future work aiming to enhance reasoning capabilities in multimodal contexts. The fully open-sourced Skywork R1V model invites further exploration and innovation within the community, potentially leading to broader applications and more advanced AI systems in the future.

This paper advances the theoretical understanding of multimodal reasoning models and offers practical insights into building efficient, high-performance AI systems. The availability of model weights for public use further underscores the commitment to reproducibility and the advancement of research.