Language models can perform multi-step reasoning using non-legible intermediate tokens.
Assessment
The evidence, examined directly, establishes the claim.
The claim says that this is possible, so one clean demonstration settles it. Such demonstrations exist. Several experiments show that language models can learn to reason step by step in human-illegible encoded formats, and this holds under reinforcement learning, supervised fine-tuning and in-context learning. In these setups the encoded intermediate tokens carry the running state of the computation, and the final answer depends on them. The most direct evidence is a 2026 study by Schulz and colleagues. It found that encoded but unconcealed reasoning is learned easily, while reasoning hidden steganographically inside innocuous cover text is much harder to learn.
Two other lines of evidence are weaker. The first is filler-token computation: models can compute over tokens that carry no content, but the gain comes from parallel work rather than a chain of serial steps. The second is that outcome-based reinforcement learning can make chains of thought illegible. Here it has not been shown that the illegible text does the work.
The claim is about capability, not about deployed behaviour. It does not mean that production models currently reason in hidden ways, and that has not been shown. It also does not mean that models reason well in obscure ciphers, which they do poorly. And it does not mean that steganographic reasoning is easy, which it is not.
Full reasoning: the evidence and decisions behind this verdict
This pass was triggered by the first assessment of the encoded-reasoning subclaim, which came back verified (confidence 0.85, credence 0.95). That subclaim is the premise of the argument from trained encoded reasoning, which is rated certain if its premise holds. The earlier evaluation held back the decisive mark only because the premise was unassessed. The premise is now established, and the subclaim's own assessment records demonstrations under RL, SFT and ICL. Schulz, Fülle and Fruengel (arxiv.org/abs/2609.39838) is the main source. Roger and Greenblatt (arxiv.org/abs/2310.18512) and Skaf et al. (2025) are earlier independent lines. The argument is therefore now marked decisive.
The step stated in the argument's prose, that the encoded tokens carry the sequential state the answer depends on, is part of what "reason step by step in encoded formats" means in those studies. It is not an independent doubtful premise.
The remaining caveats limit scope; they do not threaten the claim. Models reason poorly in lesser-known ciphers. Steganographic reasoning is hard to learn. One earlier result (Zolkowski et al.) was found to leak computation into legible channels. None of these undermines the existential claim, because the other demonstrations are separate from that one.
Confidence rises slightly from 0.78 to 0.82. It stays below 0.9 because this pass did not reread the primary experiments in full to check for leakage.
What would change the verdict: evidence that the main encoded-reasoning demonstrations work by memorized lookup or by legible leakage rather than by stepwise computation in the encoded channel.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because language models can learn to reason step by step in human-illegible encoded formats, and the encoded intermediate tokens in those demonstrations carry the running state of a sequential computation whose final answer depends on them, language models can perform multi-step reasoning using non-legible intermediate tokens.
- Does it follow?
- certain if the premises holdIf the premises are true, the conclusion must be. Whether they are is shown beside this.constitution →
- Do the premises hold?
- 1 verified
If models do learn to compute step by step in an illegible format, with the answer depending on the encoded intermediate state, the existential claim follows directly. Its premise, Language models can learn to reason step by step in human-illegible encoded formats., is now established by several independent experiments under reinforcement learning, fine-tuning and in-context learning. That makes the argument decisive. Leakage was found in one earlier study, but the other demonstrations are independent of it.
Because transformers trained with dense supervision can use meaningless filler tokens to solve algorithmic tasks they cannot solve without intermediate tokens, the extra computation that solves those multi-part problems runs over token positions whose content is uninformative to a reader, so models can reason through non-legible intermediate tokens.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting the filler-token result, it shows computation carried over tokens with no legible content. But that computation is parallel work within the network's depth, not a sequence of reasoning steps relayed through the tokens, so it bears only weakly on multi-step reasoning. Its premise, Transformers trained with dense supervision can use meaningless filler tokens to solve algorithmic tasks they cannot solve without intermediate tokens., is well documented for constructed tasks.
Because outcome-based reinforcement learning can make some reasoning models' chains of thought illegible while their answers stay correct, and penalizing specific chain-of-thought content leads models to encode that reasoning steganographically, ordinary training pressures produce models that reason through intermediate tokens a reader cannot follow.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 supported1 not separately assessed
Granting that training pressure produces illegible or encoded traces, the conclusion still needs those traces to do computational work, and neither premise establishes that. Outcome-based reinforcement learning can make some reasoning models' chains of thought illegible. is supported, but resampling found no link between legibility and accuracy. Penalizing specific chain-of-thought content leads language models to encode that reasoning steganographically. is unassessed, and in the best-known case the obfuscated traces were not shown to carry the computation.
Provenance
Where this claim has been said, linked to its canonical form.
We demonstrate a concrete failure mode where frontier models exhibit invisible reasoning by leveraging semantically irrelevant filler tokens to improve performance on synthetic reasoning tasks.
Abstract of an evaluation of 13 frontier models on filler-token benefits.
For most tasks, models learn steganographic reasoning only under SFT, while they learn steganographic messaging and encoded reasoning under all three elicitation methods.
Abstract; encoded reasoning is defined as computing step by step in an illegible but unconcealed format.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Oct 30, 2026. Every judgment on this page is accompanied by a reasoning trace.