Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning
For researchers in multimodal AI, this work reveals that the performance gains attributed to latent visual reasoning are actually due to non-visual components, highlighting the need for mechanistic evaluation beyond accuracy.
The paper decomposes latent tokens in multimodal language models into slots, boundary markers, and format, and finds that boundary markers alone preserve 78-100% of the performance gain, contradicting the visual-memory account. The gain stems from boundary markers, format, and attention patterns, not latent slots.
Recent latent visual reasoning methods achieve substantial gains by inserting continuous latent tokens into multimodal language models. These gains are commonly attributed to the tokens encoding visual evidence; recent analyses, however, reveal a paradox: the tokens are loosely tied to the image and contribute little to the answer. Critically, these analyses treat latent tokens as a single unit, obscuring the true source of the gains. We therefore decompose latent tokens into three testable components: latent slots, boundary markers, and format, and develop a state-of-the-art method as a probe under favorable conditions. Across six method-stage settings and four perception-heavy benchmarks, latent slots fail every prediction of the visual-memory account. Strikingly, retaining only the boundary markers preserves 78 to 100% of the gain in several settings, while the model attends to the image more narrowly at latent positions than at answer positions. The gain therefore comes from boundary markers, format, and this attention pattern, not from latent slots. How each method engages this mechanism depends on its training supervision: at matched accuracy, mechanisms can still differ markedly. Latent visual reasoning thus needs evaluation not only by accuracy but by what the model actually relies on.