A new paper on arXiv, "MYOPIA: Understanding the Limits of Tree-Based Reasoning in Vision-Language Models," examines a proposed way to make VLM classification more interpretable. Instead of outputting a single label, a model can follow a decision tree from root to leaf, answering a sequence of questions. The authors call this tree-based reasoning.
The approach has a clear downside. Because the tree traversal is greedy and does not backtrack, it commits to a path without revisiting earlier choices, and this is where the accuracy cost lies. The abstract frames this as a fundamental limit of tree-based reasoning, not just an implementation detail.
The paper does not claim tree-based reasoning is useless; rather, it aims to understand the trade-off between interpretability and accuracy. For researchers considering decision trees as an explanation mechanism for VLMs, the work is a caution that interpretability may come with a measurable predictive cost.