Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

M2-Reasoning: Empowering MLLMs with Unified General and Spatial Reasoning

Inclusion AI, :, Fudong Wang, Jiajia Liu, Jingdong Chen, Jun Zhou, Kaixiang Ji, Lixiang Ru, Qingpei Guo, Ruobing Zheng, Tianqi Li, Yi Yuan, Yifan Mao, Yuting Xiao, Ziping Ma

TL;DR

M2-Reasoning-7B addresses the gap in dynamic spatial reasoning for multimodal large language models by integrating a high-quality, multi-stage data pipeline with 294.2K samples and a dynamic, multi-task RLVR training framework. The approach blends a curriculum-based data ordering, step-wise optimization, and task-specific rewards to harmonize general and spatial reasoning capabilities. Empirical results across eight benchmarks show state-of-the-art performance in both general and spatial reasoning, highlighting robust, unified reasoning in diverse modalities. The work advances practical multimodal reasoning for real-world tasks by improving structured thought, temporal-spatial understanding, and instruction-following accuracy.

Abstract

Recent advancements in Multimodal Large Language Models (MLLMs), particularly through Reinforcement Learning with Verifiable Rewards (RLVR), have significantly enhanced their reasoning abilities. However, a critical gap persists: these models struggle with dynamic spatial interactions, a capability essential for real-world applications. To bridge this gap, we introduce M2-Reasoning-7B, a model designed to excel in both general and spatial reasoning. Our approach integrates two key innovations: (1) a novel data pipeline that generates 294.2K high-quality data samples (168K for cold-start fine-tuning and 126.2K for RLVR), which feature logically coherent reasoning trajectories and have undergone comprehensive assessment; and (2) a dynamic multi-task training strategy with step-wise optimization to mitigate conflicts between data, and task-specific rewards for delivering tailored incentive signals. This combination of curated data and advanced training allows M2-Reasoning-7B to set a new state-of-the-art (SOTA) across 8 benchmarks, showcasing superior performance in both general and spatial reasoning domains.

M2-Reasoning: Empowering MLLMs with Unified General and Spatial Reasoning

TL;DR

M2-Reasoning-7B addresses the gap in dynamic spatial reasoning for multimodal large language models by integrating a high-quality, multi-stage data pipeline with 294.2K samples and a dynamic, multi-task RLVR training framework. The approach blends a curriculum-based data ordering, step-wise optimization, and task-specific rewards to harmonize general and spatial reasoning capabilities. Empirical results across eight benchmarks show state-of-the-art performance in both general and spatial reasoning, highlighting robust, unified reasoning in diverse modalities. The work advances practical multimodal reasoning for real-world tasks by improving structured thought, temporal-spatial understanding, and instruction-following accuracy.

Abstract

Recent advancements in Multimodal Large Language Models (MLLMs), particularly through Reinforcement Learning with Verifiable Rewards (RLVR), have significantly enhanced their reasoning abilities. However, a critical gap persists: these models struggle with dynamic spatial interactions, a capability essential for real-world applications. To bridge this gap, we introduce M2-Reasoning-7B, a model designed to excel in both general and spatial reasoning. Our approach integrates two key innovations: (1) a novel data pipeline that generates 294.2K high-quality data samples (168K for cold-start fine-tuning and 126.2K for RLVR), which feature logically coherent reasoning trajectories and have undergone comprehensive assessment; and (2) a dynamic multi-task training strategy with step-wise optimization to mitigate conflicts between data, and task-specific rewards for delivering tailored incentive signals. This combination of curated data and advanced training allows M2-Reasoning-7B to set a new state-of-the-art (SOTA) across 8 benchmarks, showcasing superior performance in both general and spatial reasoning domains.

Paper Structure

This paper contains 26 sections, 9 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Data Construction
  3. General Reasoning Data
  4. High-Quality Multimodal Chain-of-Thought Synthesis Pipeline
  5. Difficulty Scoring for Data in Reinforcement Learning
  6. Spatial Reasoning Data
  7. Spatial Data Synthesis Pipeline
  8. Spatial Data Augmentation Pipeline
  9. Overview of Data Configuration
  10. Approach
  11. Dynamic Reasoning Incentivizing
  12. Two-Stage Training Strategy
  13. Curriculum Sampling
  14. Step-wise Dynamic Optimization
  15. Task-specific Accuracy Reward Formulation
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Benchmark performance of M2-Reasoning-7B.
  • Figure 2: Overview of the data configurations during cold-start and RLVR.
  • Figure 3: The M2-Reasoning's model architecture is built upon the Qwen2.5-7B language model, incorporating a native-resolution vision encoder. Notably, the figure omits the MLP projector typically used to connect the vision encoder and the language model.
  • Figure 4: Visualization of the EDNM reward function for different values of the hyperparameter $\lambda$.