Alibaba has unveiled a brand new artificial intelligence (AI) reasoning model that focuses on open-ended problem solving and inventive pondering. It was inspired by OpenAI’s reasoning model ‘o1’ and explained that it’s a differentiated approach from models that take care of clear and quantifiable results like existing mathematics or coding.
On the twenty first (local time), Alibaba conducted research on ‘Marco-o1’, an AI model designed to strengthen open reasoning problem-solving capabilities. Post the paper within the archivedid it
This model is a Large Reasoning Model (LRM) developed with OpenAI’s ‘o1’ as a touch. Nevertheless, while o1 performed well in math and coding benchmarks equivalent to AIME and CodeForces, Marco-o1 was designed to deal with solving unstructured tasks.
The goal is to show generalized performance in various fields where there aren’t any clear evaluation standards.
For this purpose, cutting-edge technologies equivalent to Chain-of-Thought (CoT)-based fine-tuning, Monte Carlo Tree Search (MCTS), and inference-based motion strategies were applied. These technologies help Marco-o1 solve complex problems effectively and sophisticatedly.
Particularly, CoT-based coordination, which is designed to explicitly track and manage thought processes step-by-step, makes problem solving transparent and systematic. MCTS technology explores multiple reasoning paths through the problem-solving process, and assigns a confidence rating to every path to pick probably the most promising path and derive the optimal answer.
As well as, search efficiency and accuracy are maximized by adopting a reasoning-based motion strategy that may dynamically adjust motion units when solving problems. Moreover, a self-reflection mechanism was introduced in order that the model could review and improve solutions by itself. This feature contributes to increasing accuracy in complex problems.
Marco-o1, which mixes such diverse technologies, is alleged to have the power to effectively handle not only structured tasks but in addition complex and open-ended problems.
In consequence of the benchmark test, Marco-o1 recorded an accuracy improvement of 6.17% for the English dataset and 5.60% for the Chinese dataset within the multilingual math benchmark MGSM. It also showed excellent leads to translation work, especially demonstrating high accuracy in colloquial translations that reflect cultural nuances.
“This study was inspired by OpenAI’s o1, as revealed within the model name,” the researchers said. “We aim to explore potential approaches to uncover the technology roadmap for unclear large-scale inference models.”
“The model shows similar inference characteristics to o1, and we acknowledge that performance falls wanting a completely implemented o1,” he said. “This research just isn’t a one-off, and we’re committed to ongoing optimization and continuous improvement.” added.
The present Marco-o1 model and code are hugging faceand githubOpen to the general public, anyone can use it.
Reporter Park Chan cpark@aitimes.com
