Paper 2601.19325
Innovator-VL: A Multimodal Large Language Model for Scientific Discovery
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 10
- GitHub
- 167 stars
01 In brief
Summary
Innovator-VL is a scientific multimodal large language model (MLLM) designed for scientific understanding and reasoning while maintaining general vision performance.
It uses a transparent, reproducible pipeline with RICE-ViT vision encoder, PatchMerger projector, and Qwen3-8B-Base language model.
Training includes language-image alignment (LLaVA-1.5 558k), high-quality mid-training (85M samples), supervised fine-tuning (46M samples, including scientific data like OCSR, chemical reactions, and electron micrographs), and reinforcement learning (172K samples) using Group Sequence Policy Optimization (GSPO) with a hierarchical reward system.
Innovator-VL achieves competitive results with fewer than 5 million curated scientific samples, demonstrating data efficiency.
It outperforms comparable models (7B-9B) on scientific benchmarks, especially chemistry tasks (e.g., OpenRxn 57.05%, MolParse 64.90%), and shows strong general and math reasoning performance.
The thinking model achieves state-of-the-art average score of 61.83% across 37 benchmarks.
It also exhibits superior token efficiency, using 62-66% fewer tokens than Intern-S1-mini and achieving 1.4-2x higher accuracy-to-token ratio than MiMo-VL-7B-RL.
02 From the paper
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.