Reviews
Closer readings of individual papers, and how well each claim has held up.
Several headline mechanisms in this literature have already failed replication. The useful question about a latent-reasoning result is therefore not usually whether the number is real, but whether the explanation of the number is.
Assessments rest on abstracts, reported figures, the rebuttal literature, and cross-checking between papers: not on a full reading of every PDF. Where an assessment turns on a specific number, that number was verified. Where it turns on judgement, it is a judgement and should be argued with. Corrections are welcome as issues.
Coconut read both numbers
The paper that defined the object, and it deserves that status: feeding the hidden state back instead of a token is the cleanest possible statement of what latent reasoning is, and the implementation is public and reproducible.
The difficulty is in how it gets cited. ProsQA 97.0% against 77.5% is the number everyone repeats. GSM8K 34.1% against 42.9%, a loss to ordinary chain of thought, is in the same paper and is repeated far less. Both are real, and together they say something more interesting than either alone: continuous thoughts help on search and planning, where carrying a frontier of candidates beats committing to one path, and hurt on arithmetic, where the serial discipline of writing each step down is apparently doing useful work.
Holds up: the mechanism. Zhu et al. Later proved a two-layer transformer with $D$ continuous steps solves graph reachability at diameter $D$ where the best discrete construction needs $O(n^2)$, and that the superposition emerges from training rather than being built in. Does not: any reading of this as a general improvement over chain of thought.
Let's Think Dot by Dot shines
The best-designed experiment in the literature, because it is the one that could have come out the other way. If intermediate tokens help by carrying reasoning content, meaningless ones should not help. They do, transformers solve tasks unsolvable in a single pass using nothing but repeated dots, which tells you the tokens are buying computation, not communication.
The paper is also unusually honest about its own limits: the benefit only materialises under dense, task-specific supervision, and does not emerge from ordinary training. That qualification is what separates it from the weaker filler-token claims that followed, and Vennam et al. Later confirmed the pessimistic reading for naive implementations.
The reason it matters most: filler tokens are a transcript that fully determines the model's state while revealing nothing about the computation underneath. That is a direct counterexample to the comfortable assumption that a readable transcript implies readable reasoning, and the 2026 follow-up recovering those hidden values from activations at 80–95% accuracy shows where the information actually was.
Soft Thinking mechanism did not survive
A training-free method: replace each sampled token with a probability-weighted mixture of embeddings, so a single step can carry several candidate continuations at once. Reported gains are modest but real: up to 2.48 points pass@1 with up to 22.4% fewer tokens.
The explanation did not hold. Wu et al. Showed the model overwhelmingly follows the single highest-probability component of the mixture, so the method largely reduces to greedy decoding wearing a continuous costume. Their fix, inject noise to force the exploration the original claimed was already happening, is an elegant diagnosis, since it only works if the diagnosis is right.
Why it is here: not to single out one paper, but because this is the field's characteristic failure mode. A method produces a real gain. The gain is attributed to superposition. Nobody checks whether superposition occurred. Parallel-exploration claims are per-method burdens of proof, not a property of continuous representations.
The Illusion of Superposition? shines
The paper that disciplines the subfield. It asks the obvious question nobody had asked systematically: when models are trained with continuous latent chain of thought, does the superposition everyone invokes actually appear?, and finds it does, but only in models trained from scratch. Under fine-tuning or training-free regimes, which is how nearly all practitioners work, it collapses to token commitment or shortcuts.
Consequence: the theoretical case for continuous thoughts and the practical recipes for producing them are not yet connected. The theorems are about a regime most of the applied work does not operate in. That gap is arguably the most important open problem here, and it is not primarily a theory problem.
A Formal Comparison Between Chain of Thought and Latent Thought shines
The one sharp separation in the area: continuous latent thought at polylog depth reaches $\mathsf{TC}^k$; discrete chain of thought at the same depth budget is bounded above by $\mathsf{TC}^{k-1}$. After years of informal argument that continuous states “carry more,” this is an actual theorem, though a conditional one, resting on $\mathsf{TC}^{k-1} \subsetneq \mathsf{TC}^k$, and the latent side is characterised exactly while the CoT side is only an upper bound.
What makes it a good paper rather than merely a useful one is that it supplies its own counterweight. The separation holds under deterministic decoding; move to stochastic decoding and discrete chains regain FPRAS-style approximate counting and sampling that deterministic latent thought cannot match. The authors could have shipped the headline and omitted the qualification. They did not.
Alongside: the memory-budget separation bounds it from the other side. A fixed-size recurrent state cannot match polynomial-length chain of thought on $\mathsf P$-complete problems, because the transcript's working memory grows and the loop's does not. Depth and memory are different currencies.
Recurrent depth, and Huginn-0125 shines
A 3.5B model over 800B tokens whose core block unrolls 4 to 64 times at test time, trained with no chain-of-thought data at all, and the weights, the eight intermediate checkpoints, and the full pretraining stack are public. Whatever one concludes about the approach, this is how a result should be shipped.
One framing to handle carefully: the compute-equivalence figures, often relayed as "performs like a 50B model." That is a compute-load comparison within one paper's setup, not a general scaling claim, and there is still no scaling law fitting loss jointly against parameters, loop count and tokens. The paper is more careful about this than its citations are.
Complicated by: the 2026 finding that looped performance peaks at some recurrence depth and then collapses, which means "unroll further at test time" is not a free dial and the useful range has to be found empirically.
SIM-CoT, and the instability nobody advertised shines
Its contribution is less the method than the diagnosis. As latent reasoning tokens are added, the latent vectors homogenise, lose semantic diversity, and training collapses. This had been visible in scattered form across earlier work, in curricula, staged schedules, and auxiliary losses whose necessity was never quite explained, and SIM-CoT names it.
What it implies: that end-to-end latent training under-constrains the representation, and that much of the apparatus in this literature exists to compensate. A fix requiring step-level supervision from explicit reasoning is also a quiet admission that the explicit trace is still doing the teaching.
The GPT-6 Astra system card read it narrowly
Genuinely notable: a frontier lab documenting, in its own words, "a substantial decrease in chain-of-thought monitorability compared to previous models," along with the observation that merely telling the model a monitor is present can shorten its chain of thought. That is a laboratory finding about obfuscation appearing in a deployment document.
The critique is of the surrounding discourse rather than the card. The card never names an architecture. Recurrent depth, looped transformers, opaque recurrence: none of it appears. That attribution comes from reporting. So there are two claims with different evidentiary standing, and they are being welded together: the monitorability loss is OpenAI's own documented measurement. The mechanism behind it is journalism.
Both may well be true. They are not equally established, and a literature that has just watched several mechanism claims fail replication should be the last to blur that line.
Mitigation: the card commits to "investigating alternatives to CoT monitoring (such as methods based on activation monitoring)", which is this literature's open research problem restated as a deployment roadmap.