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

Abstract: We present Innovator-VL, a scientific multimodal LLM 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.

• The paper presents Innovator-VL, a multimodal LLM that combines region-aware vision encoding, PatchMerger token compression, and chain-of-thought reinforcement learning to enhance scientific reasoning.
• The methodology employs reproducible pretraining, supervised fine-tuning, and RL across diverse datasets, achieving competitive performance in physics, chemistry, mathematics, and microscopy.
• The model demonstrates data efficiency with up to 66% fewer tokens and significant accuracy gains, paving the way for cost-effective and extensible scientific discovery.

Innovator-VL: Scientific Multimodal Reasoning and Data-Efficient Intelligence

Introduction

Innovator-VL presents a robust approach to scientific multimodal large language modeling by demonstrating that principled methodology, transparent training protocols, and curated data pipelines can deliver strong, domain-adapted intelligence without large-scale domain-specific pretraining. The model targets advanced reasoning and comprehension across physics, chemistry, mathematics, and microscopy, while preserving general computer vision capabilities and instruction-following performance. Its development directly addresses reproduction cost, extension difficulty, and generalization imbalance prevalent in prior MLLMs focused on scientific tasks.

Model Architecture

Innovator-VL adopts a modular, three-stage design integrating:

• Vision Encoder: RICE-ViT, a region-aware transformer, designed for decomposing scientific imagery into meaningful visual units necessary for downstream multimodal science reasoning. Its clustered region transformer layer achieves superior semantic segmentation and object grounding—crucial for high-fidelity parsing of structured figures, chemical diagrams, and micrographs.
• Projector: PatchMerger, which compresses dense visual token sequences into minimal but information-rich representations, significantly lowering computation and memory requirements. This architecture supports high-resolution, multi-image settings typical in scientific workflows.
• LLM: Qwen3-8B-Base, a STEM-proficient open-source LLM, enables detailed cross-modal reasoning and broad instruction following via extensive pretraining on diverse textual corpora.

Training Methodology

The methodology is fully reproducible and modular, with explicit recipes for each stage:

Pretraining

• Language-Image Alignment: A projector is aligned to word embeddings using LLaVA-1.5 558K data, establishing robust cross-modal mappings.
• Mid-Training: The model is exposed to 85M curated image-text pairs from concept-balanced corpora, employing semantic embedding strategies instead of pure brute-force text matching.

Supervised Fine-Tuning

• General Multimodal Instructions: A diverse instruction dataset enables strong performance on standard tasks.
• Chain-of-Thought and Reasoning: Honey-Data-15M drives multi-step and chain-of-thought learning, crucial for tasks involving multi-modal scientific problem-solving.
• Scientific Data Synthesis: Domain-specific protocols create paired data for (i) chemical structure/image parsing via E-SMILES format, (ii) reaction mechanism understanding from PDF-derived schemes, and (iii) microstructural electron microscopy segmentation.

Reinforcement Learning

• Fine-tuning leverages GSPO (Group Sequence Policy Optimization) for sequence-level reward alignment, emphasizing correct, concise reasoning pathways. A hierarchy of reward checks—from template adherence to symbolic and LLM-based semantic verification—ensures output quality. RL data is discrepancy-driven and standardized for optimal policy learning.

Infrastructure

Training pipelines utilize high-throughput distributed frameworks (AIAK-Training-LLM+ for Megatron-LM extension) and advanced data packing algorithms to maximize GPU utilization. Asynchronous RL integration (AReal) enables large-scale, latency-minimized policy optimization over complex reasoning datasets.

Benchmarking and Results

Evaluation: Innovator-VL was benchmarked against multiple state-of-the-art MLLMs of similar parameter scales (7B-9B) across 37 datasets, spanning general vision, mathematics, and specialized scientific exploration.

Performance Highlights

• General Vision: Innovator-VL-8B-Instruct averaged 74.5% (vs. peer SOTA Qwen3-VL-8B at 74.71%), achieving highest scores on AI2D and RealWorldQA, reflecting outstanding vision-instruction alignment.
• Math & Reasoning: RL-optimized Innovator-VL-8B-Thinking reached 55.41% (absolute +4.54% over its SFT counterpart), winning all categories against competitors—demonstrating the efficacy of sequence-level RL for multimodal chain-of-thought reasoning.
• Scientific Knowledge: Dominant results in chemistry (OpenRxn, MolParse: 57%+, 64%+) significantly outperforming all baselines (sub-17%) indicates deep robust scientific knowledge internalization, with clear gains also in molecular parsing, microstructure analysis (EMVista), protein understanding, and remote sensing.

Efficiency Analysis

Innovator-VL’s RL recipe induces compact reasoning chains, yielding 62–66% fewer tokens than Intern-S1-mini and ~2x the accuracy-to-token ratio compared to MiMo-VL-7B-RL (and 4x over Intern-S1-mini). The model’s outputs are not only more accurate but token-efficient, which is critical for latency-sensitive or resource-constrained deployments.

Implications and Future Directions

Practical Impact

The architectural and training paradigm of Innovator-VL enables domain adaptation and task extension with a substantially reduced data and compute footprint. Such approaches lower the entry barrier for scientific multimodal modeling, facilitate reproducibility, and support cost-effective deployment in diverse research and industrial workflows—spanning automated experimental design, high-throughput chemistry analysis, and microscopy-based biologic investigations.

Theoretical Consequences

Innovator-VL positions itself as evidence against the necessity of large-scale, domain-specific pretraining for strong scientific reasoning. Data efficiency, modularity, and explicit reward-driven RL optimization together form an alternative axis for scientific AGI and systematic multimodal reasoning improvement. The model demonstrates that scientific and general-purpose intelligence enhancement can coexist without mutual tradeoff.

Speculation on Future Developments

Further expansion may encompass multimodal integration of video, molecular 3D data, and temporal scientific signals; model distillation and compression for edge and mobile deployment; and interactive reasoning pipelines with external scientific databases and computational engines. Methodological transparency and modularity of Innovator-VL are well-suited to drive rapid iteration and transfer learning in foundation models for scientific discovery.

Conclusion

Innovator-VL establishes a transparent, reproducible, and data-efficient paradigm for scientific multimodal large language modeling. Its combined architectural innovations, principled training stages, and robust RL strategies result in domain-leading performance on scientific benchmarks and significant improvements in reasoning efficiency. Bridging scientific and general multimodal tasks, Innovator-VL provides a viable and extensible foundation to accelerate AI-driven research across STEM fields, supporting both rigorous theoretical inquiry and real-world application.