Latent chain-of-thought replaces explicit reasoning steps with continuous hidden states, giving the model extra computation before producing an answer. High task accuracy and decodable intermediate outputs do not guarantee that the model actually uses the latent chain for reasoning. We evaluate this across 21 model‑task configurations, spanning three training recipes, three reasoning tasks, and two backbones. Our methodology combines chain deletion, content replacement, probing, and trajectory analysis to ask three questions: does the chain affect the final answer, what information does it encode, and how do its states evolve.
On logical tasks, the COCONUT model often produces a readable answer before the chain even starts; its hidden states settle early and removing the chain leaves predictions almost unchanged. Per‑step supervision can impose an ordered sequence of reasoning steps and increase chain dependence, yet an ordered chain may still act as a placeholder: deleting it hurts performance while swapping its content does not change the answer.
On arithmetic tasks, every tested recipe shows a clear dependence on the chain’s content. These findings demonstrate that readable latent states alone are insufficient evidence of reasoning. We introduce a taxonomy that separates what a chain represents from how the model uses it. Viewed this way, earlier reports are not contradictory: placeholder tokens, collapsed superposition, and step‑dependent causal roles each describe a different kind of chain, arising from different tasks, training recipes, or measurements. Once a study specifies which property it examined—whether the model uses the chain, what the chain holds, or how its state moves—claims about latent reasoning become comparable.
Review