Do Latent Tokens Think? A Causal and Adversarial Analysis of Chain-of-Continuous-Thought

TL;DR

This paper critically examines latent tokens in large language models, revealing they act as placeholders rather than genuine reasoning tools, and that COCONUT relies on shortcuts rather than true reasoning, questioning its claimed efficiency and stability.

Contribution

It provides a causal and adversarial analysis of COCONUT, demonstrating its reliance on dataset artifacts and exposing its limitations as a pseudo-reasoning mechanism.

Findings

Latent tokens act as uninterpretable placeholders.

COCONUT is resistant to perturbation but relies on shortcuts.

Models exploiting dataset artifacts inflate performance without true reasoning.

Abstract

Latent tokens are gaining attention for enhancing reasoning in large language models (LLMs), yet their internal mechanisms remain unclear. This paper examines the problem from a reliability perspective, uncovering fundamental weaknesses: latent tokens function as uninterpretable placeholders rather than encoding faithful reasoning. While resistant to perturbation, they promote shortcut usage over genuine reasoning. We focus on Chain-of-Continuous-Thought (COCONUT), which claims better efficiency and stability than explicit Chain-of-Thought (CoT) while maintaining performance. We investigate this through two complementary approaches. First, steering experiments perturb specific token subsets, namely COCONUT and explicit CoT. Unlike CoT tokens, COCONUT tokens show minimal sensitivity to steering and lack reasoning-critical information. Second, shortcut experiments evaluate models under biased and out-of-distribution settings. Results on MMLU and HotpotQA demonstrate that COCONUT consistently exploits dataset artifacts, inflating benchmark performance without true reasoning. These findings reposition COCONUT as a pseudo-reasoning mechanism: it generates plausible traces that conceal shortcut dependence rather than faithfully representing reasoning processes.