How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision?

TL;DR

This paper provides a comprehensive analysis of latent reasoning methods, revealing their internal behaviors, issues like shortcut bias, and the impact of supervision strength on their reasoning capabilities.

Contribution

It offers the first detailed investigation into the internal mechanisms of latent reasoning, highlighting issues like shortcut behavior and the effects of supervision levels.

Findings

Latent reasoning methods often exhibit shortcut behavior, achieving high accuracy without genuine reasoning.

Latent representations can encode multiple possibilities but do not perform structured search as BFS.

Stronger supervision reduces shortcut behavior but limits hypothesis diversity.

Abstract

Latent reasoning has been recently proposed as a reasoning paradigm and performs multi-step reasoning through generating steps in the latent space instead of the textual space. This paradigm enables reasoning beyond discrete language tokens by performing multi-step computation in continuous latent spaces. Although there have been numerous studies focusing on improving the performance of latent reasoning, its internal mechanisms remain not fully investigated. In this work, we conduct a comprehensive analysis of latent reasoning methods to better understand the role and behavior of latent representation in the process. We identify two key issues across latent reasoning methods with different levels of supervision. First, we observe pervasive shortcut behavior, where they achieve high accuracy without relying on latent reasoning. Second, we examine the hypothesis that latent reasoning supports BFS-like exploration in latent space, and find that while latent representations can encode multiple possibilities, the reasoning process does not faithfully implement structured search, but instead exhibits implicit pruning and compression. Finally, our findings reveal a trade-off associated with supervision strength: stronger supervision mitigates shortcut behavior but restricts the ability of latent representations to maintain diverse hypotheses, whereas weaker supervision allows richer latent representations at the cost of increased shortcut behavior.