Wolfram Reasoning: A New Paradigm for Symbolic Mathematical Reasoning in Vision-Language Models

A research project from Georgia Tech explores enhancing the visual mathematical reasoning capabilities of Qwen3-VL using Wolfram Language, achieving improved accuracy and significantly reduced reasoning costs through GRPO reinforcement learning. Addressing the bottlenecks in mathematical reasoning for Vision-Language Models (VLMs), this study introduces the domain-specific language (Wolfram) to optimize the reasoning process, providing a new direction for AI reasoning.

Vision-Language Models face a core challenge when handling mathematical problems: how to convert visually perceived mathematical concepts into verifiable and executable reasoning processes? Traditional Python code has issues of verbosity, error-proneness, and high token consumption, leading to high reasoning costs and limited accuracy. As a domain-specific language for mathematics and symbolic computation, Wolfram Language has the advantage of concise and precise expression, making it a key choice to solve this problem.

Using Qwen3-VL-2B-Instruct as the base model, a four-stage post-training process is designed: cold-start supervised fine-tuning (establishing basic Wolfram cognition), in-context learning (guiding input-output mapping), chain-of-thought reasoning (generating intermediate steps), and GRPO reinforcement learning (Group Relative Policy Optimization). Details of GRPO include: generating 10 candidate outputs per prompt, evaluating quality via a reward model, fine-tuning parameters by injecting LoRA into attention layers, and balancing exploration and exploitation.

For the limited resources of 4 NVIDIA H200 GPUs, a series of optimizations are implemented: training acceleration (quantized LoRA to reduce memory usage, FlashAttention to optimize attention, structured pruning to remove redundancy → 3x faster training); reasoning optimization (operator fusion to reduce kernel overhead, dynamic batching for adaptive adjustment →1.5x faster reasoning). These optimizations provide reusable solutions for resource-constrained environments.

Evaluation on a subset of the ViRL39K dataset shows: Wolfram reasoning achieves a 3.33% accuracy improvement over Python reasoning, reduces reasoning token count by 75%, and has a high proportion of error-free code. Key findings include: Wolfram code is syntactically correct and directly executable, token efficiency is significantly better than Python, and there is still room for accuracy improvement (optimizable via increasing sampling count, batch size, etc.).

Experiments are based on the ViRL39K large-scale visual reasoning dataset released by TIGER-Lab. Evaluation dimensions include: the proportion of generated outputs containing Wolfram code, the proportion of code with no execution errors, the proportion of correct answers after execution by the Wolfram engine, and the average token count of prompts and outputs (including mean and standard deviation), enabling comprehensive verification of the quality and efficiency of the reasoning process.

Current limitations: 4 H200 GPUs limit the exploration of the search space, distributed training (tensor/context parallelism) needs improvement, and there is still room for accuracy optimization. Future directions: expand distributed training to break single-node limitations, increase sampling count/G value/batch size/training epochs, and deepen multimodal fusion between visual features and symbolic reasoning.

Academic contributions are based on cutting-edge research such as DeepSeek-R1 (reinforcement learning reasoning), Qwen3-VL (vision-language model), VL-Rethinker (visual reasoning reflection), Toolformer (tool usage), and QLoRA/LoRA (efficient fine-tuning). The practical significance lies in revealing the potential of domain-specific languages (DSL): compared to general-purpose languages, Wolfram has semantic precision, execution reliability, and expressive conciseness, providing new ideas for the design of AI systems in fields like mathematics.