Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search logo

Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search

Free

Empowering MLLM with o1-like Reasoning and Reflection via CoMCTS

FreeFree tier
Inputs: text, imageOutputs: text
Type
Open Source

About Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search

Mulberry is a series of Multimodal Large Language Models (MLLMs) enhanced with o1-like step-by-step reasoning and reflection capabilities, developed using Collective Monte Carlo Tree Search (CoMCTS). CoMCTS is a novel learning-to-reason method that leverages collective knowledge from multiple models to collaboratively search and identify effective reasoning paths through iterative expansion, simulation with error positioning, backpropagation, and selection. The method is used to construct the Mulberry-260k dataset, a multimodal dataset with a tree of well-defined reasoning nodes. Mulberry models demonstrate superior performance on various multimodal reasoning benchmarks, as detailed in a technical report on arXiv.

Key Features

Collective Monte Carlo Tree Search (CoMCTS) method for effective reasoning-path searching and learning
Step-by-step reasoning and reflection capabilities similar to o1
Mulberry-260k multimodal dataset with explicit reasoning nodes
Uses collective knowledge from multiple models for collaborative reasoning
Iterative operations: Expansion, Simulation and Error Positioning, Backpropagation, Selection

Pros & Cons

Pros
  • Introduces a novel collective learning approach for reasoning-path search
  • Provides explicit and well-defined reasoning steps in dataset
  • Demonstrates superiority on various multimodal benchmarks
  • Open source with code to be released
Cons
  • As a research project, practical deployment readiness may be limited
  • Requires access to multiple models for collective learning
  • Performance on real-world, diverse multimodal tasks may need further validation

Best For

Multimodal question answeringComplex reasoning tasks requiring step-by-step explanationsResearch in multimodal large language models and reasoningTraining MLLMs to reflect on and correct errors