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

Zichen Wen^*¹, Boxue Yang^*¹, Shuang Chen^¹, Yaojie Zhang^¹, Yuhang Han^¹, Junlong Ke^¹, Cong Wang^¹, Yicheng Fu^¹, Jiawang Zhao^¹, Jiangchao Yao^¹, Xi Fang^², Zhen Wang^², Henxing Cai^², Lin Yao^², Zhifeng Gao^², Yanhui Hong^², Nang Yuan^², Yixuan Li^², Guojiang Zhao^², Haoyi Tao^², Nan Wang^², Han Lyu^², Guolin Ke^², Ning Liao^³, Xiaoxing Wang^³, Kai Chen^³, Zhiyu Li^³, Feiyu Xiong^³, Sihan Hu^⁴, Kun Chen^⁴, Yanfeng Wang^¹,
Weinan E^¹†, Linfeng Zhang^²†, Linfeng Zhang^¹†

¹ School of Artificial Intelligence, Shanghai Jiao Tong University | ² DP Technology | ³ MemTensor | ⁴ Institute of Theoretical Physics, Chinese Academy of Sciences
^* Equal Contribution | ^† Corresponding Author

arXiv Code Instruct Model Thinking Model Instruct Data RL Data

Abstract

We present Innovator-VL, a scientific multimodal large language model (MLLM) designed to advance multimodal understanding and reasoning across diverse scientific domains while still maintaining excellent performance on general vision tasks. Contrary to the prevailing trend of relying on massive domain-specific pretraining data and opaque training pipelines, our work demonstrates that principled training design and transparent methodology can yield strong scientific multimodal intelligence with substantially reduced data requirements. (i) First, we provide a fully transparent and end-to-end reproducible training pipeline for scientific multimodal modeling, covering all stages from data collection and cleaning to preprocessing, supervised fine-tuning, reinforcement learning, and evaluation, together with detailed optimization and hyperparameter recipes. This enables faithful reproduction of our results and facilitates systematic extension and adaptation by the community. (ii) Second, Innovator-VL exhibits remarkable data efficiency, achieving competitive performance on a wide range of scientific tasks using fewer than five million carefully curated scientific training samples, despite not relying on large-scale scientific pretraining. These results highlight that effective scientific multimodal reasoning can be achieved through principled data selection and training strategies rather than indiscriminate data scaling. (iii) Third, Innovator-VL demonstrates strong generalization beyond scientific domains, achieving competitive performance among MLLMs of comparable size on general vision benchmarks, multimodal reasoning benchmarks, and scientific benchmarks. This indicates that scientific alignment can be integrated into a unified multimodal model without compromising general-purpose capabilities. Together, our practices suggest that, in the absence of large-scale scientific data, efficient, reproducible, and high-performing scientific multimodal models can be built, thereby providing a practical and transparent foundation for future research in scientific multimodal modeling.

Overall Illustration

Figure 1: Overall illustration of Innovator-VL.

Scientific MLLM: A unified multimodal large language model for scientific understanding and reasoning.
General + Scientific: Strong general vision–language capability without sacrificing scientific specialization.
Data-Efficient Training: Competitive scientific performance achieved with carefully curated data instead of large-scale domain-specific pretraining.
Transparent Pipeline: End-to-end reproducible training covering data curation, instruction tuning, and reinforcement learning.
Strong Generalization: Consistent performance across general vision, multimodal reasoning, and diverse scientific benchmarks.

Performance

Figure 2: Performance of Innovator-VL across general, reasoning, and scientific benchmarks.

General: Innovator-VL-8B-Instruct achieves top-tier performance on general multimodal benchmarks, matching or surpassing leading open-source models in visual perception, OCR, document understanding, and real-world reasoning.

Math & Reasoning: Innovator-VL-8B-Thinking further improves multimodal reasoning, attaining the highest average scores on visual math and reasoning benchmarks among comparable 8B models through reinforcement learning–enhanced long-horizon reasoning.

Science: Innovator-VL consistently outperforms general-purpose baselines, delivering strong results on chemistry, reaction understanding, molecular parsing, microscopy analysis, and scientific VQA tasks.

Training Data

Figure 3: Data distribution across different training stages of Innovator-VL.

Mid-Training: The mid-training stage focuses on large-scale multimodal alignment, combining diverse general-domain and scientific data to establish robust visual perception and cross-modal representation learning.

Instruction Tuning: The instruction-tuning stage emphasizes high-quality supervised data, integrating general instruction-following samples with carefully curated scientific tasks to enhance controllability and domain-specific understanding.

Reinforcement Learning: The reinforcement learning stage uses a compact but targeted dataset to further refine long-horizon reasoning, response consistency, and decision-making quality.

Data Construction Pipeline

Figure 4: Overview of data construction pipelines for scientific multimodal training.

Optical Chemical Structure Recognition (OCSR): A human-in-the-loop pipeline combines large-scale synthetic bootstrapping with active-learning-driven expansion on real patent and paper data, using E-SMILES as a unified annotation format to represent complex chemical structures.

Chemical Reaction Understanding: Reaction datasets are constructed from scientific PDFs via automated layout parsing and expert-verified question generation, covering both fine-grained reaction perception and document-level multimodal reasoning.

Electron Microscopy (EM) Microstructures: Large-scale EM datasets are built through iterative expert annotation and model-assisted refinement, with dense instance-level segmentation and structured attribute descriptions to support microstructural analysis.

Scientific Data Overview

Curated multimodal scientific data supporting reproducible and data-efficient training

Total Scientific Samples

≈4.8M

Scientific Domains

Modalities

Domain-wise Composition

Scientific Domain	Primary Tasks	Modalities
Chemistry	OCSR, reaction understanding, molecule parsing	Image + Structured Text
Materials Science	Microstructure analysis, EM interpretation	Microscopy Image + Text
Scientific Documents	Figure understanding, document-level reasoning	PDF Layout + Image + Text
General Science QA	Visual question answering, reasoning	Image + Natural Language

Training Stage Statistics

Stage	Dataset	Scale	Purpose
Mid-Training	Innovator-VL-Mid-Training	85M samples	Multimodal alignment and representation learning
Instruction Tuning	Innovator-VL-Instruct	46M samples	Instruction following and scientific controllability
Reinforcement Learning	Innovator-VL-RL	172K trajectories	Long-horizon reasoning and decision refinement

On Token Efficiency of Reasoning

Figure 5: Accuracy and token efficiency comparison on mathematical reasoning benchmarks.

Innovator-VL exhibits strong efficiency in reasoning token consumption. As shown in Figure 5, Innovator-VL-8B-Thinking achieves competitive or superior accuracy on mathematical reasoning benchmarks while using substantially fewer tokens than comparable baselines, leading to lower inference cost and reduced latency.

This efficiency is primarily driven by the reinforcement learning stage, which encourages the model to focus on critical reasoning steps and eliminate redundant computation, resulting in more compact and effective reasoning trajectories.

BibTeX

@article{wen2026innovator,
  title={Innovator-VL: A Multimodal Large Language Model for Scientific Discovery},
  author={Wen, Zichen and Yang, Boxue and Chen, Shuang and Zhang, Yaojie and Han, Yuhang and Ke, Junlong and Wang, Cong and others},
  journal={arXiv preprint arXiv:2601.19325},
  year={2026}
}