The AI industry was once a quiet, niche space, rarely making waves beyond its own sphere.
That all changed when OpenAI launched ChatGPT. Since then, tech companies big and small have been racing to create cutting-edge AI tools to address a wide range of needs and applications.
While many innovations have emerged from the West, the East is proving it’s not far behind in this competitive landscape.
Alibaba is one of the most prominent tech companies from China, and here, its researchers revealed what is supposed to be a direct competitor to OpenAI-o1, a specialized Large Language Model capable of reasoning.
Alibaba calls its own take the 'Marco-o1'.
According to its Hugging Face page and GitHub page:
The idea is to be an AI capable of solving complex problems classic language models often struggle with.
More or less, Marco-o1 is designed to be just like OpenAI-o1
In fact, the researchers did say that they were inspired by OpenAI-o1, saying that:
But Marco-o1 is more than just a OpenAI-o1 copy because the researchers said that it leverages advanced techniques that can distinguish itself.
Read: Alibaba Introduces MIMO, A Way To Create A Controllable Character Through AI Video Synthesis
Highlights include:
- Fine-Tuning with CoT Data: The researchers created Marco-o1-CoT by performing full-parameter fine-tuning on the base model using open-source CoT dataset combined with their own self-developed synthetic data.
- Solution Space Expansion via MCTS: The researchers integrated LLMs with MCTS (Marco-o1-MCTS), using the model's output confidence to guide the search and expand the solution space.
- Reasoning Action Strategy: The AI implements novel reasoning action strategies and a reflection mechanism (Marco-o1-MCTS Mini-Step), including exploring different action granularities within the MCTS framework and prompting the model to self-reflect, thereby significantly enhancing the model's ability to solve complex problems.
- Application in Translation Tasks: The researchers claim that Marco-o1 is the first to apply Large Reasoning Models (LRM) to Machine Translation task, exploring inference time scaling laws in the multilingual and translation domain.
This allows it to have an exceptionally advanced ability in handling complex tasks that traditional LLMs may have difficulties with.
"MCTS allows exploration of multiple reasoning paths using confidence scores derived from softmax-applied log probabilities of the top-k alternative tokens, guiding the model to optimal solutions. Moreover, our reasoning action strategy involves varying the granularity of actions within steps and mini-steps to optimize search efficiency and accuracy," the researchers said.