TLDR: M2-Reasoning-7B is a new Multimodal Large Language Model (MLLM) that significantly improves both general and spatial reasoning. It achieves this through a novel data pipeline that generates high-quality training data and a dynamic multi-task training strategy with tailored rewards. The model sets new state-of-the-art records across 8 benchmarks, demonstrating enhanced capabilities in understanding complex problems and dynamic spatial interactions, while also acknowledging areas for future improvement.
Recent advancements in Multimodal Large Language Models (MLLMs) have significantly boosted their reasoning capabilities, especially with techniques like Reinforcement Learning with Verifiable Rewards (RLVR). However, these models have often struggled with understanding dynamic spatial interactions, a crucial skill for real-world applications.
To address this challenge, researchers have introduced M2-Reasoning-7B, a new model designed to excel in both general problem-solving and spatial understanding. This innovation is built upon two core components: a unique data pipeline and a dynamic multi-task training strategy.
A Novel Data Approach
The first key innovation is a sophisticated data pipeline that generates a massive 294.2K high-quality data samples. These samples are divided into two sets: 168K for initial ‘cold-start’ fine-tuning and 126.2K for the RLVR stage. What makes this data special is its focus on logically coherent reasoning trajectories, ensuring that the model learns from well-structured thought processes. The data undergoes a comprehensive assessment to guarantee its quality, difficulty, and diversity, which is vital for effective learning.
For general reasoning, the pipeline synthesizes high-quality multimodal chain-of-thought data, filtering it based on answer accuracy and detailed reasoning quality. It also includes a prompt difficulty scoring method for RLVR, allowing the model to learn progressively from easier to more complex tasks.
For spatial reasoning, a dedicated data synthesis pipeline creates high-quality, semantically meaningful data from controlled spatial simulations. This includes image-based tasks (like object counting, spatial relations, and distances) and video-based tasks (such as room size, appearance order, and relative direction). The data is further enhanced through augmentation strategies that diversify questions, options, and instructions, ensuring robust training.
Dynamic Training for Unified Reasoning
The second innovation is a dynamic multi-task training strategy with step-wise optimization. This approach helps mitigate conflicts that arise from data heterogeneity and delivers tailored incentive signals through task-specific rewards. The training process involves two stages:
- Cold-start: Supervised fine-tuning on a large dataset activates the model’s latent reasoning capabilities and standardizes its output format.
- Dynamic Multi-task RLVR: Reinforcement Learning with Verifiable Rewards is applied to data with verifiable answers. This stage encourages the model to adopt correct reasoning processes and improve generalization across diverse multimodal tasks. The model uses a variant of GRPO (Generalized Reinforcement Learning with Policy Optimization) with dynamic hyper-parameter adjustments and a curriculum sampling approach, where training data is organized by increasing difficulty.
The reward system is also finely tuned. For general reasoning, a rule-based mechanism evaluates exact matches for multiple-choice and fill-in-the-blank questions. For spatial reasoning, where exact numerical matches can be challenging, an Exponential Decay Numeric Matching (EDNM) reward function provides a smoother, continuous reward based on normalized relative error, encouraging the model to optimize in the correct direction even with initial inaccuracies.
Also Read:
- Bridging Language and Logic: A New Framework for AI Reasoning
- Leanabell-Prover-V2: Advancing Formal Theorem Proving with AI Self-Correction
Setting New Performance Standards
M2-Reasoning-7B has been rigorously evaluated across eight distinct benchmarks, demonstrating its superior performance in both general and spatial reasoning domains. In general reasoning, it achieved a new state-of-the-art average score of 45.0, outperforming other leading models. It particularly excelled in benchmarks like MathVista and DynaMath.
For spatial reasoning, M2-Reasoning-7B also set a new state-of-the-art on CV-Bench with an average score of 82.3, showing exceptional strength in understanding complex spatial configurations, relations, depth, and distance. On the more challenging VSI-Bench for nuanced video spatial imagination, it demonstrated highly competitive performance, establishing new records for inferring Room Size and determining Relative Direction.
While M2-Reasoning-7B marks a significant leap forward, the researchers acknowledge ongoing challenges, including constrained reasoning depth, occasional pathological repetition in generated responses, and areas for improvement in fine-grained visual perception. Future work aims to address these limitations to further enhance the model’s robustness and reasoning capabilities. For more technical details, you can refer to the full research paper: M2-Reasoning: Empowering MLLMs with Unified General and Spatial Reasoning.
