Beyond Hidden Chain-of-Thought: Mechanistic Evidence for Distributed Latent Reasoning

Latent-reasoning models perform multi-step inference in hidden-state space, offering a more efficient alternative to explicit chain-of-thought (CoT), but leave open what computation their internal states implement. We study CODI, a continuous chain-of-thought model that reasons through a sequence of latent states, and ask whether these latent states compress explicit reasoning or instead implement a distinct computational process. Using causal interventions and probes, including latent-step ablations, hidden-state injection, activation patching, and wrong-prefix probing, we test whether latent thoughts function as compressed versions of explicit reasoning steps. Our results show that latent reasoning exhibits computational properties similar to chain-of-thought reasoning, including multiple successful trajectories and gradual refinement across steps, but we do not find evidence that its internal states form a hidden chain-of-thought. Instead, answer-relevant information is distributed across latent iterations and transformer layers, with only partial compatibility between explicit CoT representations and latent computation. In the wrong-prefix setting, the model faithfully encodes perturbed intermediate values in the residual stream, yet we find no evidence that latent states explicitly encode whether those values are correct or incorrect. Despite this finding, latent reasoning is substantially more robust than explicit chain-of-thought reasoning under identical perturbations. Overall, our findings suggest that CODI implements a distributed reasoning process related to, but not directly interchangeable with, explicit chain-of-thought reasoning.