Language models can perform multi-step reasoning using non-legible intermediate tokens.
5 events · 2 assessments · 2 decisions
Reassessed
The encoded-reasoning subclaim's first assessment came back verified (0.85). The argument from trained encoded reasoning is now marked decisive. Status stays verified and confidence rises from 0.78 to 0.82. The other arguments were reconfirmed as slight. Dependents were not notified because the status did not change.
Reassessed: still Verified
verdict confidence 0.78 → 0.82 · credence 0.93
Structured and assessed
First pass. Three for-arguments: trained encoded reasoning (new subclaim 0fe2726c), filler tokens (new subclaim 5ffae9ff), emergent illegibility (linked existing d8eba1e3 and caca00ca; the Matcher matched a fine-tuning capability wording to caca00ca, which I judged a near neighbour and used as a supporting subclaim while minting the broader learnability claim separately after a fresh match returned new). Assessed verified (0.78), credence 0.92, on Schulz et al. 2026, Roger & Greenblatt 2023, Pfau et al. 2024, Baherwani et al. 2026. Importance 0.45, contestation 0.4. Parent b06f66a2 already describes this claim consistently (capability in constructed settings), so no notification needed beyond the new assessment being visible; notifying it anyway as first assessment.
Assessed Verified
verdict confidence 0.78 · credence 0.92
Read as an existential capability claim, that language models can be made to carry out multi-step reasoning through intermediate tokens a reader cannot follow, the claim is established by several independent demonstrations. The clearest come from work on encoded reasoning: models can learn to reason step by step in human-illegible encoded formats, whether fine-tuned to do so (Roger and Greenblatt, 2023), trained by reinforcement learning against word-level monitors (Skaf et al., 2025), or given a scheme in context. A 2026 study by Schulz, Fülle and Fruengel found this kind of illegible but unconcealed reasoning was learned readily under reinforcement learning, supervised fine-tuning and in-context learning alike. Filler-token experiments add a second line: transformers trained with dense supervision can solve otherwise unsolvable algorithmic tasks using meaningless filler tokens (Pfau, Merrill and Bowman, 2024), and a 2026 evaluation reports frontier models gaining up to 13 points on synthetic tasks from filler tokens. A third line is suggestive but not conclusive: outcome-based reinforcement learning can make some models' traces illegible, and penalizing chain-of-thought content may push reasoning into encoded form. The capability has clear limits, and these are where the live questions sit. Filler tokens only help on problems that can be parallelized within a single forward pass, not on inherently serial ones, so they show hidden computation rather than long chains of hidden steps. Steganographic reasoning, meaning reasoning hidden inside text that looks like it is about something else, is much harder to learn than plain encoded reasoning. Schulz and colleagues found it learned mostly under supervised fine-tuning, and readily only on a cover task that makes hiding easy. No production model has been shown to reason steganographically. In current deployed models, illegible stretches of trace have not been clearly shown to do computational work. So "can" is well supported for constructed and trained settings. How far the capability reaches in deployed models doing long serial reasoning is the open part.
Claim entered the graph