TLDR: WE-MATH 2.0 is a new system designed to enhance the mathematical reasoning abilities of AI models, especially with visual problems. It features a structured mathematical knowledge system (491 knowledge points, 1,819 principles), two new datasets (MathBook-Standard and MathBook-Pro with 3D difficulty modeling), and a two-stage reinforcement learning training method (MathBook-RL). The system also introduces MathBookEval, a comprehensive benchmark. Experiments show significant improvements in generalization and robustness, demonstrating effective learning with limited data and better handling of complex, multi-step problems.
A new research paper introduces WE-MATH 2.0, a comprehensive system designed to significantly improve how Multimodal Large Language Models (MLLMs) handle complex mathematical reasoning, especially when visual information is involved. While MLLMs have shown impressive abilities in various tasks, they often struggle with the nuances of mathematical problem-solving.
The researchers behind WE-MATH 2.0 identified several key challenges in existing approaches. These include a lack of a comprehensive system for mathematical knowledge, difficulty in modeling problem complexity from a model’s perspective, and a tendency for models to memorize problems rather than generalize their reasoning skills. To tackle these issues, WE-MATH 2.0 integrates a structured mathematical knowledge system, a unique way of modeling data difficulty, and a training method based on reinforcement learning.
The Core Components of WE-MATH 2.0
The system is built on four main contributions:
1. MathBook Knowledge System: This is a meticulously organized, five-level hierarchical system that covers 491 distinct mathematical knowledge points and 1,819 fundamental principles. This structure, derived from sources like Wikipedia and textbooks and refined by human experts, provides a systematic way to supervise mathematical learning for MLLMs.
2. MathBook-Standard & Pro Datasets: MathBook-Standard is a dataset designed for broad conceptual coverage and flexibility. It uses a ‘dual expansion’ strategy, meaning it includes multiple images for a single question and multiple questions for a single image, enriching visual and semantic diversity. Building on this, MathBook-Pro introduces a three-dimensional difficulty space, modeling ‘step complexity’ (number of knowledge points), ‘visual complexity’ (added auxiliary elements in images), and ‘contextual complexity’ (linguistic scenario variations). Each problem in MathBook-Pro has seven progressive difficulty variants, enabling structured and gradual learning for MLLMs. Notably, all images in these datasets are handcrafted using GeoGebra software, ensuring precision and rigor.
3. MathBook-RL Training Paradigm: This is a two-stage reinforcement learning framework. The first stage, ‘Cold-Start Fine-tuning,’ teaches the MLLM to reason in a knowledge-oriented, step-by-step manner. The second stage, ‘Progressive Alignment RL,’ uses a curriculum-based approach with dynamic data scheduling. This stage helps the model progressively align its reasoning across different difficulty levels, improving generalization and robustness.
4. MathBookEval Benchmark: To thoroughly assess MLLMs’ reasoning capabilities, MathBookEval was developed. This benchmark covers all 491 knowledge points with diverse reasoning step distributions, providing a comprehensive tool for evaluating how well models understand and apply mathematical concepts.
Also Read:
- Evaluating Continuous Learning in Multimodal AI: Introducing MLLM-CTBench
- Putnam-AXIOM: A New Benchmark Reveals LLM Mathematical Reasoning Gaps
Experimental Findings
Experiments show that MathBook-RL performs very well compared to existing models on four widely-used mathematical reasoning benchmarks. It significantly improves performance over its base model, Qwen2.5-VL-7B, by over 5% across all benchmarks. The progressive alignment reinforcement learning proved particularly effective in enhancing knowledge generalization, especially on tasks requiring multi-step reasoning.
Interestingly, the system achieves strong performance using a relatively small amount of training data (only 9.8K samples). This efficiency is attributed to the high-quality, structured mathematical knowledge system, which allows for effective alignment and generalization even with limited data.
Further analysis on MathBookEval revealed that MLLMs’ performance decreases as the number of required knowledge points increases, especially for problems needing 7-10 knowledge points. Models also performed better in algebra than in geometry, highlighting ongoing challenges in spatial reasoning. Larger models generally showed more consistent improvements across all difficulty levels and knowledge domains.
The research also explored the impact of the fine-tuning stage. While supervised fine-tuning alone offered limited gains, it was crucial for unlocking the full potential of reinforcement learning. Additionally, using natural language for chain-of-thought reasoning during fine-tuning proved more effective than structured step-wise formats, suggesting that flexible reasoning prompts are beneficial.
WE-MATH 2.0 represents a significant step forward in developing more capable and generalizable MLLMs for visual mathematical reasoning. The project’s resources, including the datasets and GeoGebra files, will be made publicly available, fostering further research and potentially aiding in mathematics education. You can find the full research paper here: WE-MATH 2.0 Research Paper.
