Virgo: A Preliminary Exploration on Reproducing o1-like MLLM
FreeReproducing o1-like slow-thinking reasoning in multimodal LLMs
FreeFree tier
Inputs: text, imageOutputs: text
About Virgo: A Preliminary Exploration on Reproducing o1-like MLLM
Virgo (Visual reasoning with long thought) is a multimodal slow-thinking system that fine-tunes a capable multimodal large language model (MLLM) with a small amount of textual long-form thought data. It is a preliminary exploration into reproducing o1-like reasoning capabilities in MLLMs, demonstrating that slow-thinking capacities are fundamentally associated with the language model component and can be transferred across modalities or domains. The research finds that textual long-form reasoning data can be even more effective than visual reasoning data for eliciting slow-thinking capacities in MLLMs. Resources are released as open source.
Key Features
Fine-tunes capable MLLMs with textual long-form thought data to enable slow visual reasoning
Demonstrates that slow-thinking capacities transfer across modalities (text to vision)
Shows textual reasoning data can be more effective than visual reasoning data for eliciting slow thinking
Preliminary open-source technical report with released resources
Built upon existing MLLM frameworks
Pros & Cons
Pros
- Open-source and free to use
- Novel approach leveraging textual thought data for multimodal reasoning
- Effectively transfers slow-thinking capacities from text to vision
- Provides a foundation for further research on multimodal slow thinking
Cons
- Preliminary exploration with limited evaluation scale
- Not peer-reviewed (arXiv preprint)
- Relies on existing capable MLLMs, not a standalone model
- Performance on complex real-world tasks not fully assessed
Best For
Multimodal slow-thinking reasoning tasksVisual question answering requiring deep reasoningAcademic research on o1-like reasoning in MLLMsExploring modality transfer of reasoning capabilities
FAQ
What is Virgo?
Virgo (Visual reasoning with long thought) is a multimodal slow-thinking system that fine-tunes a capable MLLM with textual long-form thought data to reproduce o1-like reasoning in multimodal contexts.
How does Virgo achieve slow thinking?
It fine-tunes a capable MLLM on a small amount of textual long-form thought data, showing that slow-thinking capacities are primarily linked to the language model component and can transfer across modalities.
Is Virgo open source?
Yes, the resources are released as open source, and the paper is available on arXiv with code and data.
What makes Virgo different from other MLLMs?
It specifically targets slow-thinking reasoning inspired by o1-like systems, using only textual thought data to elicit multimodal slow thinking, which is a novel approach.