Latent reasoning: research directions

Each node is a paper or a hub concept, coloured by literature. Edges are labelled by relation: builds on, same object read two ways, endpoint of the dial, rebuttal, and not yet connected. The last kind joins works that were searched for and not found connected as of 16 September 2026, and each such gap is listed with its confidence under the map.

Corrections welcome: file an issue on microprediction/latentreasoning.

Gaps

Gap edges are claims about the literature, so each carries a confidence. Each is a dated, scoped search claim rather than a proof of absence, and is withdrawn when the missing work is found.

1. Interpretability results have not reached the deployment question. The finding that gold traces can be recovered from latent tokens on most correctly-answered arithmetic instances, and the controversy over a production model whose card records reduced monitorability do not cite each other. Nobody has applied the decoding methods to the deployed case. Confidence: high. The most policy-relevant gap here.
2. No open-checkpoint measurement of monitorability loss. An external capability evaluation is attached. UK AISI ran four non-agentic tests on Astra and found a no-CoT reasoning horizon of 30.9 minutes against 3.6 for the previous model, a constraint-following rate of 93% against 48%, a more compressed reasoning style, and missing reasoning summaries on up to 80% of long simulated cyber trajectories. The card then states that “UK AISI did not directly test if Astra evades monitors successfully and makes no claims about the overall monitorability of the model,” and that agentic monitorability evaluations could not be completed. The separate finding that awareness of a monitor shortens the chain of thought comes from OpenAI's own adversarial honeypot evaluation, not from AISI. So what is missing is an external test of monitor evasion, and a shared public suite run on open recurrent-depth checkpoints. Confidence: high.
3. No shared benchmark for latent computation. Narrowed. A purpose-built instrument does exist: a 4,000-item benchmark that forces internal inference by requiring the answer to be signalled through the language of the first response token, so it cannot be externalised as text. What is still missing is community convergence: it measures internal reasoning in stock models rather than discriminating between latent-reasoning architectures, its authors cannot rule out heuristic exploitation, and critique papers continue to build bespoke diagnostics instead of adopting a common one. Confidence: moderate, for the narrowed claim.
4. Which auxiliary objective, and why. A common claim holds that policy-gradient objectives assume single-step decisions and so cannot train latent recurrence. That is false: for a differentiable recurrence with gradients retained, a terminal reward's total derivative already includes every unrolled update. What is true is narrower: standard objectives have empirically underperformed on looped models, and the causes on offer (long Jacobian products, sparse reward, truncation or detaching in implementations) are optimisation problems, not invalidity. The open question is which auxiliary objectives and truncation/halting choices actually improve gradient quality and compute efficiency under controlled comparison. Confidence: moderate, and now a question rather than an absence.
5. The superposition arguments still stand apart. Narrowed. Padding and looping are now analysed together with threshold-circuit characterisations, so the broad claim that the toolkits never meet was wrong. What has not been connected is the superposition-and-parallel-trace line of argument to either of them. Confidence: moderate.
6. The mixture-of-experts and latent-reasoning literatures do not read each other. A mixture-of-experts gate reads a token's hidden state and picks an action, which is latent routing by any definition, and it is the most deployed instance of it anywhere. Three dedicated latent-reasoning surveys from 2025 cite no mixture-of-experts routing work, the dedicated mixture-of-experts survey cites no latent reasoning or looped transformers, and the flagship papers on each side do not mention each other. Mixture-of-Recursions is the single bridge, repurposing expert-choice routing to select recursion depth and citing both. Confidence: high, checked by full-text search of the four surveys.
7. Weights-as-memory and cache-as-memory have been compared only on long-context language modelling. TTT-E2E writes the test context into updated weights and is measured against full and sliding-window attention, which keeps the context in the cache. Full attention wins on needle-in-a-haystack recall. No paper runs that comparison on multi-step reasoning. KaVa, the nearest cache-side reasoning method, compresses a teacher's chain of thought during training, not a test context. Confidence: medium.
8. State-space recursion is untested beyond tiny scale. Narrowed. Mamba-2 hybrid operators have been placed inside a latent recursive scaffold, so the blanket claim was false. But that work is ~7M parameters on abstract reasoning tasks, and nothing establishes a linear-state model as the reasoning medium at language-model scale. Confidence: medium, for the narrowed claim only.
9. No agreed compute-matched protocol for comparing latent reasoners against chain of thought. Every paper picks its own basis. One scaling study compares at fixed nominal parameter count with loop count varied, a retrofitting study compares a 6M adapter against a 180M block and against a “same-size” scratchpad model, and a third compares against unconverted baselines with no stated parameter or FLOP match. The clearest evidence is a 2025 efficiency benchmark built, in its own words, because “fair comparison of efficiency-oriented approaches is hindered by fragmented evaluation practices.” Confidence: high.
10. No hop-count sweep on naturalistic tasks at frontier scale. Controlled studies do sweep hop count: Yao et al. train GPT-2-scale models from scratch on $k$-hop datasets for $k = 2, 3, 4$ and find data requirements growing exponentially and depth linearly in $k$. Those use synthetic templated data and models up to 770M. The work on pretrained frontier models stops at two hops, where it finds a scaling trend for the first hop but not the second. What is missing is the sweep on naturalistic knowledge tasks in pretrained models, with matched shortcut controls and mechanistic interventions. Confidence: moderate, for the narrowed claim.
11. The headline continuous-thought result has no independent replication. Searches for a failed or matched-compute replication of the founding continuous-thought paper return nothing in either direction, and the same holds for the recursion-depth routing method. The nearest available substitute is a mechanistic study of a descendant method finding it does not execute a full latent rollout at longer hop lengths. Confidence: medium, and an absence of checking rather than a failure of the result.

Found a counterexample to any of these? That is the most useful possible contribution: open an issue. The four withdrawn above arrived that way.