Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO
Authors: Huanjin Yao, Qixiang Yin, Jingyi Zhang, Min Yang, Yibo Wang, Wenhao Wu, Fei Su, Li Shen, Minghui Qiu, Dacheng Tao, Jiaxing Huang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations over 6 widely-used reasoning benchmarks showcase the superior performance of our method. ... In this section, we first provide implementation details in Sec. 4.1, and then present main results in Sec. 4.2 that demonstrate the effectiveness of Share-GRPO. In Sec. 4.3, we conduct comprehensive ablation studies to examine the impact of each design in Share-GRPO. |
| Researcher Affiliation | Collaboration | 1Nanyang Technological University 2Byte Dance 3Tsinghua University 4Beijing University of Posts and Telecommunications 5The University of Sydney 6 Sun Yat-sen University |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. The methodology is described in continuous text and mathematical formulas. |
| Open Source Code | Yes | Code is available at https://github.com/HJYao00/R1-Share VL. |
| Open Datasets | Yes | For training data, we randomly sample 52K multimodal data from MM-Eureka [5]. ... We provide an extensive comparison against state-of-the-art models across 6 widely used and challenging benchmarks, covering a diverse range of reasoning tasks from specialized domains to general-purpose reasoning. [62, 63, 64, 65, 66, 67] |
| Dataset Splits | Yes | For training data, we randomly sample 52K multimodal data from MM-Eureka [5]. ... We provide an extensive comparison against state-of-the-art models across 6 widely used and challenging benchmarks, covering a diverse range of reasoning tasks from specialized domains to general-purpose reasoning. |
| Hardware Specification | Yes | Model optimization is carried out using Easy R1 [53] codebase, with training conducted on 8 NVIDIA H100 GPUs for the 7B model and 32 H100 GPUs for the 32B model. |
| Software Dependencies | No | Model optimization is carried out using Easy R1 [53] codebase, with training conducted on 8 NVIDIA H100 GPUs for the 7B model and 32 H100 GPUs for the 32B model. No specific version numbers for software libraries are listed. |
| Experiment Setup | Yes | For RL related hyperparameters, we use a global batch size of 128, a rollout batch size of 512, a rollout temperature of 0.7, and a learning rate of 1e-6. |