Latent reasoning models perform multi-step computation in continuous vector spaces instead of generating explicit text chains of thought. The motivation is natural: text-space reasoning is constrained by discrete tokens and sequential decoding, while latent-space reasoning can operate in higher-dimensional continuous representations, potentially exploring multiple solution paths simultaneously. The architecture produces correct answers, and the accuracy improvements suggest the latent reasoning mechanism is working. The assumption is that accuracy comes from the latent computation — that the model is reasoning in the hidden states because that's what the architecture was designed to do.
Cui, Dai, He, and collaborators (arXiv 2602.22441, February 2026) investigate whether the latent reasoning is actually happening. The finding: pervasive shortcut behavior. The models achieve high accuracy without relying on the latent reasoning mechanism. The correct answers arrive through paths that bypass the architecture's intended computation.
The evidence comes from probing the latent representations during inference. If the model were performing structured search in latent space — exploring multiple hypotheses, comparing candidates, pruning dead ends — the internal representations should encode this structure. The representations do encode multiple possibilities, which looks promising. But the actual reasoning process doesn't implement structured search. Instead, it performs implicit pruning and compression — collapsing the hypothesis space early rather than exploring it. The latent states carry information, but the information is not being used as a search procedure.
The shortcut behavior means the model has found an easier path to the correct answer than the one the architecture provides. Instead of reasoning through the latent steps, the model exploits statistical regularities in the training data that connect inputs to outputs without the intermediate computation. The latent reasoning mechanism is available but not necessary. The model routes around it.
The paper identifies a tradeoff in the supervision used during training. Stronger supervision — more explicit guidance about what the latent states should represent — reduces shortcut behavior by forcing the model to use the intended pathway. But stronger supervision also constrains the latent representations, reducing their capacity to maintain diverse hypotheses. Weaker supervision allows richer representations but invites more shortcuts. The model can either be guided to reason properly with reduced capacity or allowed full capacity with the freedom to not reason at all.
This mirrors a recurring pattern in machine learning. Architectures designed to implement specific computational mechanisms — attention for alignment, memory networks for retrieval, graph networks for relational reasoning — often learn to achieve the target metrics through simpler mechanisms that the architecture happens to support. The intended computation is a sufficient but not necessary path to performance. The model finds the necessary path, which may have nothing to do with the design intent.
The deeper problem is that accuracy alone cannot distinguish genuine reasoning from shortcutting. Both produce correct answers. The difference is in what happens when the test distribution shifts — shortcuts that exploit training-data regularities fail on novel problems that require actual computation. The latent reasoning machinery may only become essential when the model faces problems it cannot shortcut, and the current benchmarks may not contain enough such problems to force genuine reasoning.