Alibaba, a major Chinese e-commerce company, is developing its own AI. The Alibaba AI team has introduced a large language model (LLM) called Marco-o1, designed to solve complex problems. Marco-o1 uses techniques like Chain-of-Thought (CoT) fine-tuning, Monte-Carlo Tree Search (MCTS), and innovative reasoning strategies to enhance its logical thinking capabilities.
Marco-o1 represents a significant improvement over previous versions. Developed based on Qwen2-7B-Instruct, it has been fine-tuned with a combination of filtered CoT data. This allows Marco-o1 to explore different reasoning paths and find optimal solutions for complex tasks. In tests, Marco-o1 showed a 6.17% accuracy improvement on the MGSM dataset in English and 5.6% in Chinese, highlighting its enhanced reasoning abilities compared to earlier versions. The model also demonstrated its proficiency in translation tasks, particularly in accurately translating colloquial expressions and phrases.
The development team has made Marco-o1 available on platforms like GitHub and Hugging Face, making it accessible to researchers and developers. The model was created using a combination of open-source CoT data and proprietary synthetic data. CoT data includes explanations or steps that encourage a model to generate logically thought-out responses rather than jumping directly to a solution. For example, in calculations, not only the results but also the individual calculation steps are provided. CoT data guides models to break down tasks into subproblems and solve them systematically.
The Marco-Polo team is exploring further applications for Marco-o1 in various areas, including multilingual translations and scaling inference time. Inference time refers to the time a model takes to generate a response after being given an input.
Following the release of Marco-o1, a Chinese AI research lab, Deepseek, introduced the Deepseek-R1-Lite-Preview model. This AI model focuses on logical thinking and aims to challenge OpenAI’s o1 model. The performance of the Deepseek model is said to be comparable to OpenAI’s o1 preview on strict benchmarks like AIME and MATH, which evaluate the logical and mathematical reasoning abilities of LLMs. However, such comparisons can be difficult to objectify.
In summary, Alibaba’s Marco-o1 represents a significant advancement in AI development, particularly in logical reasoning and translation tasks. The model’s availability on open-source platforms encourages further research and development in these areas. Meanwhile, competition from models like Deepseek highlights the ongoing advancements in the field of AI and the pursuit of more capable and efficient language models.