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

In a new arXiv paper, researcher Karthic Palaniappan explores a novel approach to vision-language model (VLM) reasoning by incentivizing neuro-symbolic language-based thinking via reinforcement learning (RL). Taking inspiration from the 2016 film Arrival, where learning an alien language grants the ability to transcend time, the study aims to teach VLMs to represent and reason with concepts in a structured, symbolic language rather than natural language alone. The base model used is Qwen3-VL-2B-Instruct, trained on 4× Nvidia H200 GPU nodes.

The results are striking: the neuro-symbolic RL method improved accuracy by 3.33% on a vision-language evaluation dataset spanning math, science, and general knowledge questions, while reducing the number of reasoning tokens by 75% over the symbolic math library SymPy. This suggests that encoding reasoning steps in a compact, symbolic form can make VLMs both more accurate and far more efficient. Palaniappan also documents compute challenges, scaling possibilities, and future work to improve thinking in neuro-symbolic languages for VLMs, with the full training and inference setup available on GitHub.