MarcoPolo Team Unveils Marco-o1 Reasoning Model for Complex Real-World Problem-Solving and Open-Ended Tasks

The MarcoPolo team has recently released Marco-o1, a large-scale reasoning model that specializes in solving complex real-world problems and answering open-ended questions. This model has demonstrated exceptional performance in the field of multi-language translation, achieving accuracy increases of 6.17% on the MGSM English dataset and 5.60% on the MGSM Chinese dataset.

Marco-o1's reasoning capabilities have been enhanced through the use of thought chain fine-tuning, MCTS (Monte Carlo Tree Search) to expand the search space, and innovative reasoning strategies. Furthermore, the model has shown outstanding performance on both standardized and open-ended tasks, outperforming other models in its class.

Marco-o1 is based on the Qwen2-7B-Instruct model and has been open-sourced on both Hugging Face and GitHub, providing a valuable resource for the development of future reasoning models.