Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Innovator-VL: A Multimodal Large Language Model for Scientific Discovery

Zichen Wen, Boxue Yang, Shuang Chen, Yaojie Zhang, Yuhang Han, Junlong Ke, Cong Wang, Yicheng Fu, Jiawang Zhao, Jiangchao Yao, Xi Fang, Zhen Wang, Henxing Cai, Lin Yao, Zhifeng Gao, Yanhui Hong, Nang Yuan, Yixuan Li, Guojiang Zhao, Haoyi Tao, Nan Wang, Han Lyu, Guolin Ke, Ning Liao, Xiaoxing Wang, Kai Chen, Zhiyu Li, Feiyu Xiong, Sihan Hu, Kun Chen, Yanfeng Wang, Weinan E, Linfeng Zhang, Linfeng Zhang

TL;DR

Innovator-VL introduces a transparent, data-efficient scientific multimodal LLM that preserves general vision capabilities while delivering strong cross-domain reasoning. The architecture combines a region-aware vision encoder (RICE-ViT), a learned PatchMerger projector, and a pre-trained LLM backbone (Qwen3-8B-Base), trained through a two-stage pre-training process and staged post-training that includes SFT and reinforcement learning with a hierarchical reward. It achieves competitive general performance and state-of-the-art results on scientific benchmarks, notably in chemical reaction understanding, while using fewer than 5 million curated samples. The work emphasizes reproducibility and principled data usage, and demonstrates notable token efficiency gains through RL-driven reasoning, suggesting practical implications for scalable, trustworthy scientific AI.

Abstract

We present Innovator-VL, a scientific multimodal large language model designed to advance understanding and reasoning across diverse scientific domains while maintaining excellent performance on general vision tasks. Contrary to the trend of relying on massive domain-specific pretraining and opaque pipelines, our work demonstrates that principled training design and transparent methodology can yield strong scientific intelligence with substantially reduced data requirements. (i) First, we provide a fully transparent, end-to-end reproducible training pipeline, covering data collection, cleaning, preprocessing, supervised fine-tuning, reinforcement learning, and evaluation, along with detailed optimization recipes. This facilitates systematic extension by the community. (ii) Second, Innovator-VL exhibits remarkable data efficiency, achieving competitive performance on various scientific tasks using fewer than five million curated samples without large-scale pretraining. These results highlight that effective reasoning can be achieved through principled data selection rather than indiscriminate scaling. (iii) Third, Innovator-VL demonstrates strong generalization, achieving competitive performance on general vision, multimodal reasoning, and scientific benchmarks. This indicates that scientific alignment can be integrated into a unified model without compromising general-purpose capabilities. Our practices suggest that efficient, reproducible, and high-performing scientific multimodal models can be built even without large-scale data, providing a practical foundation for future research.

Innovator-VL: A Multimodal Large Language Model for Scientific Discovery

TL;DR

Innovator-VL introduces a transparent, data-efficient scientific multimodal LLM that preserves general vision capabilities while delivering strong cross-domain reasoning. The architecture combines a region-aware vision encoder (RICE-ViT), a learned PatchMerger projector, and a pre-trained LLM backbone (Qwen3-8B-Base), trained through a two-stage pre-training process and staged post-training that includes SFT and reinforcement learning with a hierarchical reward. It achieves competitive general performance and state-of-the-art results on scientific benchmarks, notably in chemical reaction understanding, while using fewer than 5 million curated samples. The work emphasizes reproducibility and principled data usage, and demonstrates notable token efficiency gains through RL-driven reasoning, suggesting practical implications for scalable, trustworthy scientific AI.

Abstract

We present Innovator-VL, a scientific multimodal large language model designed to advance understanding and reasoning across diverse scientific domains while maintaining excellent performance on general vision tasks. Contrary to the trend of relying on massive domain-specific pretraining and opaque pipelines, our work demonstrates that principled training design and transparent methodology can yield strong scientific intelligence with substantially reduced data requirements. (i) First, we provide a fully transparent, end-to-end reproducible training pipeline, covering data collection, cleaning, preprocessing, supervised fine-tuning, reinforcement learning, and evaluation, along with detailed optimization recipes. This facilitates systematic extension by the community. (ii) Second, Innovator-VL exhibits remarkable data efficiency, achieving competitive performance on various scientific tasks using fewer than five million curated samples without large-scale pretraining. These results highlight that effective reasoning can be achieved through principled data selection rather than indiscriminate scaling. (iii) Third, Innovator-VL demonstrates strong generalization, achieving competitive performance on general vision, multimodal reasoning, and scientific benchmarks. This indicates that scientific alignment can be integrated into a unified model without compromising general-purpose capabilities. Our practices suggest that efficient, reproducible, and high-performing scientific multimodal models can be built even without large-scale data, providing a practical foundation for future research.
Paper Structure (37 sections, 12 equations, 5 figures, 2 tables)

This paper contains 37 sections, 12 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Model Architecture
  3. Vision Encoder
  4. Projector
  5. Language Model
  6. Pre-training
  7. Post-training
  8. Supervised Fine-tuning
  9. General Multimodal Instruction Data
  10. Chain-of-Thought and Multi-step Reasoning Data
  11. Scientific Understanding Data
  12. Reinforcement Learning
  13. RL Data
  14. Discrepancy-Driven Selection and Filtering.
  15. Standardization and Composition.
  16. ...and 22 more sections

Figures (5)

  • Figure 1: Innovator-VL-8B performance across different benchmarks. The first row displays the results of the Innovator-VL-8B-Instruct model on general benchmarks, while the second and third rows present the Innovator-VL-8B-Thinking model on mathematical reasoning and scientific benchmarks. The yellow dashed line represents the average score of the selected models.
  • Figure 2: Overall architecture of our native-resolution multi-image reasoning model. Given a text prompt and multiple images with heterogeneous resolutions, RICE-ViT encodes each image at its native size, producing variable-length visual token sequences. Patch Merger then aggregates the visual tokens into a compact sequence that is concatenated with text tokens and fed into Qwen-8B-base to generate the final response with explicit reasoning.
  • Figure 3: Data distribution across different training stages. (a) Distribution of data sources within the Innovator-VL-Mid-Training-85M dataset. (b) Distribution of data sources within the Innovator-VL-Instruct-46M dataset. (c) Distribution of data sources within the Innovator-VL-RL-172K dataset.
  • Figure 4: Data Construction Pipeline. The raw data are obtained from both synthetic generation and real-world sources. For each data modality (EM representations, images, and questions), domain experts apply modality-specific inspection and refinement strategies. Through repeated iterative optimization, a final high-quality dataset is produced.
  • Figure 5: Token efficiency comparison across vision reasoning benchmarks. (a) Average token lengths, showing that Innovator-VL-8B-Thinking generates significantly shorter reasoning chains compared to the other models. (b) Accuracy-to-token ratio, which measures the reasoning efficiency, demonstrating that Innovator-VL-8B-Thinking achieves $1.4\times$ to $2\times$ higher accuracy-to-token ratio than MiMo-VL-7B-RL and $3.9\times$ to $4.3\times$ higher than Intern-S1-mini.