Alibaba Introduces Marco-o1 to Rival OpenAI’s o1

The MarcoPolo Team, part of Alibaba International Digital Commerce, has launched Marco-o1, an advanced large language model (LLM) built for addressing reasoning needs in open-ended problem-solving tasks.

Powered by techniques such as Chain-of-Thought (CoT) fine-tuning, Monte Carlo Tree Search (MCTS), and novel reasoning strategies, Marco-o1 seeks to push the boundaries of AI models in handling complex, real-world challenges.

The development team has made Marco-o1 available on GitHub and Hugging Face, making it accessible for researchers and developers. The model is based on the Qwen2-7B-Instruct architecture and has been fine-tuned using a combination of filtered Open-O1 CoT dataset, Marco-o1 CoT dataset, and Marco-o1 Instruction dataset.

The model was developed using a combination of open-source CoT data and proprietary synthetic data. Key features include MCTS, which allows the model to explore multiple reasoning paths using confidence scores, and the integration of reasoning action strategies that refine the model's problem-solving approach through varying action granularities.

In its current implementation, Marco-o1 has shown a 6.17% improvement in accuracy on the MGSM English dataset and 5.60% on the Chinese version, demonstrating its enhanced reasoning power. 

The model also excels in machine translation tasks, accurately translating complex phrases and slang expressions, such as converting a literal translation of “这个鞋拥有踩屎感” to “This shoe has a comfortable sole.”

The MarcoPolo Team continues to explore ways to apply Marco-o1 across multiple domains, including multilingual translation and inference time scaling.

This follows DeepSeek, a Chinese AI research lab backed by High-Flyer Capital Management, releasing the DeepSeek-R1-Lite-Preview, a reasoning AI model intended to challenge OpenAI’s o1 model.

The model’s performance is reportedly on par with OpenAI’s o1-preview on rigorous benchmarks such as AIME and MATH, which evaluate the logical and mathematical reasoning skills of LLMs.