What's Holding Back Latent Visual Reasoning?

TL;DR

This paper investigates why latent visual reasoning models often ignore intermediate tokens, revealing issues with dataset informativeness and token prediction quality that hinder their effectiveness.

Contribution

The study uncovers that current models do not utilize latent tokens effectively due to dataset limitations and token quality, providing insights for future improvements.

Findings

Latent tokens are often ignored because they add limited information in existing datasets.

Models can rely on latent tokens when trained on diagnostic datasets with informative intermediate steps.

Inference-time latent tokens tend to collapse, reducing their usefulness for reasoning.

Abstract

Humans can approach complex visual problems by mentally simulating intermediate visual steps, rather than reasoning through language alone. Inspired by this, several works on Vision-Language Models have recently explored chain-of-thought reasoning with continuous latent tokens as intermediate visual imagination steps. In this work, we investigate how recent models leverage such latent tokens. Surprisingly, we find that model accuracy is unaffected when latent tokens are replaced by uninformative dummy tokens. This indicates that latent tokens play a minimal causal role in the model's final prediction. To better understand this phenomenon, we analyze both the training signal provided by oracle latent representations and the quality of the latent tokens generated at inference time. Our experiments reveal two crucial issues holding back latent visual reasoning: First, in most existing datasets, oracle latent tokens provide limited additional information beyond the original image and do not substantially simplify the task, leading models to ignore them during training and effectively bypassing them at inference time. When fine-tuned on a diagnostic dataset, in which latent tokens provide sufficient support for the final prediction, we show that models can causally rely on them. Second, the latent tokens produced at inference time deviate from their corresponding oracle representations, collapsing to a narrow region and preventing benefits even when the model relies on them. Overall, our findings suggest that future progress in latent visual reasoning depends on two key pillars: high-quality datasets with informative intermediate steps and more precise latent token prediction.