Incentivizing Neuro-symbolic Language-based Reasoning in VLMs via Reinforcement Learning

arXiv cs.CL / 4/27/2026

📰 NewsDeveloper Stack & InfrastructureIdeas & Deep AnalysisModels & Research

Key Points

  • The paper proposes a neuro-symbolic approach that uses language-based reasoning within vision-language models (VLMs) and evaluates improvements in analytical reasoning performance and efficiency.
  • Using Qwen3-VL-2B-Instruct with a reinforcement-learning setup on 4× NVIDIA H200 GPU nodes, the authors report a 3.33% accuracy gain on a vision-language benchmark covering math, science, and general knowledge.
  • The method is also reported to significantly cut reasoning compute costs by reducing reasoning tokens by 75% compared with a SymPy-based baseline.
  • The authors discuss practical compute/scaling challenges and share the training and inference setup via a public repository for reproducibility and future work.

Abstract

There are 7,407 languages in the world. But, what about the languages that are not there in the world? Are humans so narrow minded that we don't care about the languages aliens communicate in? Aliens are humans too! In the 2016 movie Arrival, Amy Adams plays a linguist, Dr. Louise Banks who, by learning to think in an alien language (Heptapod) formed of non-sequential sentences, gains the ability to transcend time and look into the future. In this work, I aim to explore the representation and reasoning of vision-language concepts in a neuro-symbolic language, and study improvement in analytical reasoning abilities and efficiency of "thinking systems". With Qwen3-VL-2B-Instruct as base model and 4 \times Nvidia H200 GPU nodes, I achieve an accuracy improvement of 3.33\% on a vision-language evaluation dataset consisting of math, science, and general knowledge questions, while reducing the reasoning tokens by 75\% over SymPy. I've documented the compute challenges faced, scaling possibilities, and the future work to improve thinking in a neuro-symbolic language in vision-language models. The training and inference setup can be found here: https://github.com/i-like-bfs-and-dfs/wolfram-reasoning.