Scientific Multimodal Large Language Model for Advanced Reasoning

---

📰 News

• 2026.1.29 🤗🤗 Our model has been officially released at the "Agentic Science at Scale" workshop!
• 2026.1.29 🤗🤗 We’ve already adapted it to SciMaster. Come and give it a try!
• 2026.1.28 🤗🤗 We release the Innovator-VL Technical Report!
• 2026.1.26 🤗🤗 We release all of our code, model, data, and pipeline! Check out the InnovatorVL collection on Hugging Face

📖 Table of Contents

• Introduction
• Models & Checkpoints
• Datasets
• Architecture
• Performance
• Training Pipeline
• Quick Start
• Evaluation
• Directory Structure
• Citation
• Acknowledgments

🎯 Introduction

Innovator-VL is a scientific multimodal large language model designed to advance multimodal understanding and reasoning across diverse scientific domains. Contrary to conventional approaches that rely on massive scientific domain-specific pretraining, Innovator-VL demonstrates remarkable data efficiency, achieving competitive performance using fewer than five million carefully curated scientific samples. Meanwhile, 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.

Key Highlights

🚀 Superior Performance: State-of-the-art results on scientific benchmarks with 8B parameter model

💡 Data Efficiency: Competitive performance without large-scale scientific pretraining

🔬 Strong Generalization: Maintains excellent performance on general vision and reasoning tasks

📊 Fully Transparent: Reproducible training methodology from data collection to evaluation

📦 Models & Checkpoints

📊 Datasets

Training Datasets

🏗️ Architecture

Innovator-VL adopts a principled architecture design optimized for scientific understanding:

Visual Encoder: RICE-ViT captures fine-grained, region-level semantics for accurate perception of structured visual elements (symbols, annotations, relational components)

Vision-Language Projector: PatchMerger balances representational capacity and computational efficiency by merging visual patches into compact yet semantically informative representations

Language Decoder: Qwen3-8B-Base provides a strong foundation for reasoning and generation, pre-trained on a broad and diverse corpus

📊 Performance

🚀 Training Pipeline

Stage 1: Alignment

• Data: LLaVA-558K alignment dataset
• Training: Vision adapter parameters only
• Goal: Align visual and textual representations

Stage 1.5: Mid-Training

• Data: 85M high-quality multimodal samples
• Training: Full model
• Goal: Enhance multimodal understanding

Stage 2: Instruct Tuning

• Data: 46M instruction data
• Training: Full model
• Goal: Develop instruction-following capabilities

RL: GSPO Optimization

• Algorithm: Group Sequence Policy Optimization
• Data: 172K discrepancy-driven RL samples
• Goal: Enhancing model reasoning and the ability to solve complex scientific problems

Inference with HuggingFace

from transformers import AutoTokenizer, AutoProcessor, AutoModelForCausalLM
from qwen_vl_utils import process_vision_info
model_path = "InnovatorLab/Innovator-VL-8B-Instruct"

# default: Load the model on the available device(s)
model = AutoModelForCausalLM.from_pretrained(
    model_path, torch_dtype="auto", device_map="auto", trust_remote_code=True
)

# default processer
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

⚡ Quick Start

Installation

# Clone repository
git clone https://github.com/InnovatorLab/Innovator-VL.git
cd Innovator-VL

# Install dependencies
pip install -r requirements.txt

Model Conversion

# HF → Megatron
bash examples/innovator_vl/convert/convert_8b_hf_to_mcore.sh

# Megatron → HF
bash examples/innovator_vl/convert/convert_8b_mcore_to_hf.sh

Stage 1 Training

# Download LLaVA-558K dataset
# Set environment variables and run
AIAK_TRAINING_PATH=/path/to/Innovator-VL \
DATA_PATH=/path/to/LLaVA-558K \
TOKENIZER_PATH=/path/to/Innovator-VL-8B-stage0 \
CHECKPOINT_PATH=/path/to/Innovator-VL-8B-stage0 \
bash examples/innovator_vl/quick_start/train_auto_stage_1_alignment_innovator_vl_8b.sh

Stage 1.5 Training

# Download Mid-Training dataset (85M samples)
AIAK_TRAINING_PATH=/path/to/Innovator-VL \
DATA_PATH=/path/to/LLaVA-OneVision-1.5-Mid-Training-85M \
TOKENIZER_PATH=/path/to/Innovator-VL-8B-stage0 \
CHECKPOINT_PATH=/path/to/stage_1_output \
bash examples/innovator_vl/quick_start/train_auto_stage_1.5_mid_training_innovator_vl_8b.sh

Stage 2 Training

# Download instruct dataset (46M samples)
AIAK_TRAINING_PATH=/path/to/Innovator-VL \
DATA_PATH=/path/to/Innovator-VL-Instruct-46M \
TOKENIZER_PATH=/path/to/Innovator-VL-8B-stage0 \
CHECKPOINT_PATH=/path/to/stage_1.5_output \
bash examples/innovator_vl/quick_start/train_auto_stage_2_instruct_innovator_vl_8b.sh

RL Training (GSPO)

cd RL/train_scripts
# Configure paths in configs/innovator-vl-8b-gspo.yaml
bash run_example.sh

🔬 Evaluation

Innovator-VL is evaluated using the lmms-eval framework, enabling standardized, reproducible evaluation across general vision, mathematical reasoning, and scientific domains.

We provide an official evaluation script that covers all supported benchmarks.

📜 Evaluation Script

An official evaluation script is provided:

👉 lmms-eval/eval_innovator_vl.sh

Before running the evaluation, please set the model path:

export MODEL_PATH=/path/to/Innovator-VL-8B-Instruct (or Innovator-VL-8B-Thinking)

Then launch the evaluation:

cd lmms-eval
bash eval_innovator_vl.sh

🧪 Supported Benchmarks

The evaluation script is organized into three major categories.

🌍 General Vision Benchmarks

These benchmarks evaluate perception, OCR, document understanding, and real-world multimodal reasoning:

• AI2D: ai2d, ai2d_no_mask
• OCRBench
• ChartQA
• MMMU: mmmu_val, mmmu_pro_standard
• MMStar
• VStar-Bench
• MMBench: mmbench_en_dev, mmbench_en_test
• MME-RealWorld: mmerealworld, mmerealworld_cn
• DocVQA: docvqa_val
• InfoVQA: infovqa_val
• SEEDBench: seedbench, seedbench_2_plus
• RealWorldQA

➗ Math & Reasoning Benchmarks

These benchmarks focus on mathematical reasoning and chain-of-thought multimodal problem solving:

• MathVision: mathvision_reason_test_reasoning, mathvision_reason_testmini_reasoning
• MathVerse: mathverse_testmini_reasoning
• MathVista: mathvista_testmini_cot_reasoning
• WeMath: wemath_testmini_reasoning

🔬 Scientific Benchmarks

These benchmarks evaluate Innovator-VL’s capabilities across chemistry, biology, medicine, earth science, and scientific document understanding:

• General Science & Scientific Cognition: sfe-en, sfe-zh, scienceqa
• Chemistry & Molecular Science: rxnbench_vqa, OpenRxn, MolParse, SmolInstruct, superchem, superchem_cn
• Biology, Medicine & Microscopy: EMVista, microvqa, ProteinLMBench
• Earth Science & Remote Sensing: msearth_mcq, xlrs-lite

📁 Directory Structure

Innovator-VL/
├── examples/innovator_vl/          # Training examples and scripts
│   ├── quick_start/                # Training scripts for all stages
│   ├── convert/                    # Model conversion utilities
│   └── evaluate/                   # Evaluation scripts
│
├── aiak_training_llm/              # Core training framework
│   ├── models/innovator_vl/        # Model implementations
│   │   ├── innovator_vl_model.py
│   │   ├── innovator_vl_config.py
│   │   ├── adapter.py
│   │   └── vision_model.py
│   └── train.py                    # Main training script
│
├── tools/                          # Utilities
│   └── convert_checkpoint/         # Checkpoint conversion tools
│       └── custom/innovator_vl/    # Model-specific converters
│
├── RL/                             # RL training framework
│   ├── train_scripts/              # RL training scripts
│   ├── configs/                    # RL configurations (GSPO)
│   ├── trains/                     # RL training core
│   ├── engine/                     # RL training engine
│   ├── reward/                     # Reward functions
│   └── 3rdparty/                   # Third-party dependencies
│
├── ds/                             # Custom training implementations
│   └── innovator_vl/               # Model definitions
│       ├── configuration_innovator_vl.py
│       └── modeling_innovator_vl.py
│
├── asset/                          # Assets (logos, figures)
│   ├── logo.png
│   ├── architecture.pdf
│   ├── performance.pdf
│   └── data_pipeline.pdf
│
├── requirements.txt                
├── LICENSE                         
└── README.md                      

📝 Citation

If you find Innovator-VL helpful for your research, please consider citing our technical report:

@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}
}

🙏 Acknowledgments

We express our sincere gratitude to the open-source community for their invaluable contributions that made this work possible:

• LLaVA-OneVision-1.5: Fully Open Framework for Democratized Multimodal Training
• RICE-ViT: Advanced visual encoder for fine-grained region understanding
• Qwen3: Excellent language model backbone
• AReaL: A Large-Scale Asynchronous Reinforcement Learning Freamwork
• lmms-eval: A standardized evaluation framework for Large Multimodal Models
• Megatron-LM: Efficient, scalable training for large language models
• Bee: A High-Quality Corpus and Full-Stack Suite to Unlock Advanced Fully Open MLLMs
• OpenMMReasoner: Pushing the Frontiers for Multimodal Reasoning with an Open and General Recipe
• SGLang: SGLang is a fast serving framework for large language models and vision language models
• vLLM: A high-throughput and memory-efficient inference and serving engine for LLMs

These projects have significantly influenced our work, and we are deeply grateful to their respective authors and contributors.