Ariadne: A Controllable Framework for Probing and Extending VLM Reasoning Boundaries
This work addresses the problem of extending vision-language models' reasoning capabilities for spatial tasks, which is important for applications like navigation and robotics, though it's incremental as it focuses specifically on post-training methods.
The researchers investigated whether reinforcement learning post-training can extend vision-language models' capability boundaries for visual-centric spatial tasks, where base models initially fail completely. Using their Ariadne framework with synthetic mazes and difficulty-aware curriculum training, they achieved over 50% accuracy on problems where the base model scored 0%, and demonstrated zero-shot improvements of 16% on MapBench and 24% on ReasonMap benchmarks.
While Vision-Language Models (VLMs) post-trained with Reinforcement Learning (RL) show impressive general reasoning, their evaluation is often confined to language-dominant tasks (e.g., math). This raises a critical question: can RL post-training truly extend the inherent capability boundary of a base VLM, particularly for visual-centric spatial tasks where it initially fails? To investigate this, we introduce Ariadne, a framework utilizing synthetic mazes for multi-step spatial reasoning where task difficulty (e.g., path length, turns) is precisely controlled. We leverage this controllable environment to train VLMs using Reinforcement Learning with Verified Rewards (RLVR) in a difficulty-aware curriculum. Surprisingly, post-RLVR training, the VLM achieves over 50% accuracy on a problem set where the base model scored 0%, demonstrating that our approach expands the model's initial capability boundary. To assess real-world viability, we evaluate out-of-distribution (OOD) generalization on practical benchmarks. Despite training only on synthetic maze samples, Ariadne achieves significant zero-shot improvements, averaging 16% on MapBench (e.g., museum navigation) and 24% on ReasonMap (subway transfer tasks). These results confirm that our method not only broadens the model's fundamental limits but also enhances its generalization to real-world spatial reasoning. We acknowledge our study is limited to the post-training phase, given the opaqueness of pre-training data, and hope our research motivates further work on specialized, capability-extending alignment.