TLDR: Ariadne is a new framework that uses synthetic mazes and Reinforcement Learning with Verified Rewards (RLVR) to significantly improve Vision-Language Models’ (VLMs) spatial reasoning capabilities. It enables VLMs to achieve over 50% accuracy on tasks where they previously failed, and demonstrates strong zero-shot generalization to real-world navigation benchmarks like MapBench and ReasonMap, with average improvements of 16% and 24% respectively. The framework provides a controllable environment to probe and extend VLM reasoning boundaries, showing that RL post-training can broaden a model’s fundamental limits and enhance its ability to handle complex visual-centric spatial tasks.
Vision-Language Models (VLMs) have shown remarkable progress in understanding and generating content across both visual and linguistic domains. However, a significant challenge remains: their ability to perform complex visual-centric spatial reasoning, especially in tasks where they initially struggle. Traditional evaluations often focus on language-heavy tasks, leaving a gap in understanding how well these models can truly navigate and reason in visual environments.
Addressing this, a new research paper introduces Ariadne, a controllable framework designed to investigate and expand the reasoning capabilities of VLMs. The core idea behind Ariadne is to use synthetic mazes as a controlled environment to train VLMs. These mazes allow researchers to precisely adjust task difficulty, such as path length and the number of turns, providing a clear way to test and improve a model’s spatial reasoning.
The framework employs Reinforcement Learning with Verified Rewards (RLVR) within a difficulty-aware curriculum. This means the models are trained using a system that provides clear feedback on their performance, guiding them to learn more effective navigation strategies. The results are quite compelling: after RLVR training with Ariadne, a VLM that initially scored 0% on a specific problem set achieved over 50% accuracy. This demonstrates a significant expansion of the model’s inherent capability boundary.
Beyond synthetic environments, the researchers also evaluated Ariadne’s real-world applicability. Despite being trained exclusively on synthetic maze samples, the framework showed impressive zero-shot generalization on practical benchmarks. On MapBench, which involves tasks like museum navigation, Ariadne achieved an average improvement of 16%. For ReasonMap, which focuses on subway transfer tasks, the improvement was even more substantial, averaging 24%. These findings confirm that the method not only enhances the model’s fundamental limits but also improves its ability to generalize to real-world spatial reasoning challenges.
The methodology behind Ariadne leverages Group Relative Policy Optimization (GRPO), a learning framework that compares groups of candidate responses to optimize the model’s policy without needing a separate reward model. A custom reward function was designed to measure stepwise correctness, assigning proportional rewards for both fully and partially correct answers, scaled by the number of reasoning turns. This encourages the model to generate accurate and structured reasoning steps.
For training, the Qwen2.5-VL-7B-Instruct model was used as the base VLM, trained with 4,700 samples from the AlphaMaze dataset. The dataset’s difficulty was controlled by varying maze step counts and turns, ensuring exposure to both simpler and more complex scenarios. This approach helped the model acquire stable low-level navigation patterns and learn global spatial reasoning for coherent path planning.
The paper highlights that while current VLM reasoning strategies like Chain of Thought (CoT) prompting and Supervised Fine-Tuning (SFT) improve performance, they often fall short on complex tasks. Ariadne’s RL-based approach, particularly with verifiable, rule-based rewards, offers a robust way to enhance reasoning, especially in multi-step spatial understanding. The research acknowledges that its study is limited to the post-training phase, given the opaqueness of pre-training data, and hopes to motivate further work on specialized, capability-extending alignment.
Also Read:
- GeoFM: A New Approach to Boost Geometric Reasoning in AI Models
- Diagnosing AI’s Reasoning Abilities with TempoBench
In conclusion, Ariadne presents a powerful framework for extending VLM reasoning boundaries, particularly in visual-spatial tasks. By using controllable synthetic environments and RLVR, it significantly boosts model performance and demonstrates strong generalization to real-world navigation challenges. This work paves the way for more capable and robust VLMs in complex reasoning scenarios. You can read the full research paper here.
