Latent Reasoning

Research on language models that reason in continuous space instead of words.

A language model that reasons step by step writes each step down. The step is rounded to the nearest token, emitted, and read back in. That rounding is not incidental: it is a quantiser sitting in the middle of the reasoning loop, and it throws away almost everything the model computed. Latent reasoning is the study of what happens when you take it out.

The question has two halves that pull against each other. Removing the quantiser widens the channel between reasoning steps by three orders of magnitude, which buys parallel search, shorter traces, and computation that no vocabulary can express. It also removes the bound that currently forces a model to route serial reasoning through text, which is not the same as saying the text was ever a faithful account of the reasoning, and the evidence that it was not is substantial.

One axis: how many bits of each reasoning step survive to the transcript. A convenient way to arrange the work below: not a claim that the field organises itself this way.

A transcript that determines a computation is a long way from a transcript that explains it. The introduction works through the mechanism and its limits.

The state of play, September 2026

Fifteen months separate the naming of continuous thought from four dedicated survey papers and a production controversy. That velocity is itself evidence the question is real. Whether the answer has arrived is a different matter, and it depends on which claim is being made.

The efficiency claims mostly hold up, but "efficiency" needs splitting

Depth-recurrent and latent methods beat token-matched explicit chain of thought on reported accuracy per inference FLOP and on thought-phase latency, and open weights exist at 1.4B to 3.5B parameters. But parameter savings, training compute, inference compute and wall-clock latency are four different ledgers, and a win in one is not a win in the others: the iso-depth scaling study prices a recurrence at roughly $r^{0.46}$ equivalent unique parameters and finds a training-compute cost to parameter sharing in its setup. Claims here should say which ledger they are drawn on.

The claims about how the model reasons are contested, and weakening

The theoretical selling point of continuous thoughts is superposition: holding several reasoning paths at once rather than committing to one. That mechanism is proven for a specific construction and demonstrated in models trained from scratch. It was also shown not to occur in one prominent method that claimed it, and shown to collapse under the fine-tuning regimes practitioners actually use. Recurrent depth does not scale monotonically either: performance peaks and then falls with further looping.

Continuous and discrete steps separate, conditionally

Continuous latent steps reach $\mathsf{TC}^k$ where discrete steps at matched depth reach only $\mathsf{TC}^{k-1}$: under deterministic decoding. Under stochastic decoding, discrete chains win back counting and sampling power. They are complementary resources rather than a ranking, which is an unsatisfying answer and appears to be the true one. The theory page has the statements and their premises.

The interesting region is nearly empty

Almost all work sits at one endpoint or the other. Hybrids that interleave latent and textual steps are the thinnest part of the literature and the part where the efficiency and oversight arguments could actually be traded against each other rather than asserted.

Deployment

In September 2026 the question stopped being architectural. OpenAI released GPT-6 Astra and its system card records “a substantial decrease in chain-of-thought monitorability compared to previous models”: a frontier lab documenting the loss as a measured property of a shipped model. The card does not name an architecture. The attribution to recurrent depth comes from reporting, and the two claims should be kept apart. The position paper that anticipated all of this was published thirteen months earlier by forty authors drawn from the labs now on both sides of it.

The honest version of that argument cuts in both directions, and the monitorability page sets it out with the measurements attached: chains of thought were never as faithful as the safety case implies, and losing an unreliable channel before its replacement is ready is still a loss.