Outcome-based reinforcement learning can make some reasoning models' chains of thought illegible.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The claim is existential and modest: it says outcome-based RL can degrade chain-of-thought legibility in some models, not that it always does. Three lines of evidence point the same way. DeepSeek's own report states that R1-Zero, trained from a base model with outcome rewards alone, produced poorly readable, language-mixed reasoning. DeepSeek also reports that language consistency fell over RL training unless a consistency reward was added, which ties the change to the RL process itself. Arun Jose's study of 14 reasoning models (arXiv 2510.27338) found that R1, R1-Zero and QwQ often emit illegible stretches that their base model does not, while their final answers stay readable. Separately, OpenAI and Apollo Research documented idiosyncratic jargon in o3's reasoning traces.
The evidence also limits how far the claim reaches. Claude models stayed largely legible in Jose's study, so the effect depends on training details. Jose found that truncating models to only their legible portions cut accuracy, but resampling showed no correlation between legibility and performance, so whether the illegible text does real work is still unsettled. What remains contested is how general the effect is, what causes it (vestigial tokens, compressed encodings, or training artifacts), and whether it grows with scale. The narrow claim that it can happen is well supported.
Full reasoning: the evidence and decisions behind this verdict
Prior model: before reading, the claim seemed very likely true. DeepSeek-R1-Zero's readability problems are well known, and pure-RL training from a base model is about as clean a causal setup as deployed models offer. The evidence matches that expectation.
Evidence weighed: - DeepSeek-R1 report (arXiv 2501.12948; Nature 2025) describes R1-Zero as having poor readability and language mixing, and reports an ablation in which language consistency declined during RL without a language-consistency reward. This is the strongest causal evidence, because the starting model and the training signal are known. It was recalled from the report rather than re-read in this pass. It comes from the model's developer, but the developer had no motive to overstate a weakness. - Jose, arXiv 2510.27338 (abstract read): legibility was graded across 14 models; RL "often causes reasoning to become illegible", every model family except Claude did this, legibility degraded on harder questions, and accuracy fell 53% when models were forced to use only the legible portions. The study relies on one author's LLM-grader setup and on cross-model comparison. It names its own caveats (training artifacts, vestigial tokens). - The o3 jargon documented in OpenAI/Apollo anti-scheming work is consistent with the claim but was not re-verified in this pass.
Caveat: "language mixing" and "illegible" overlap without being identical. Jose's grading measures illegibility to humans and monitors directly, which closes most of that gap.
What would change the verdict: evidence that the illegibility in these models comes from something other than outcome RL (for example a decoding or sampling artifact, or SFT data), and that RL alone keeps CoT legible. The R1-Zero ablation makes this unlikely.
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.
DeepSeek-R1-Zero was trained from a base model with outcome rewards alone, and its chains of thought show poor readability and language mixing; within that training run, language consistency declines as RL proceeds unless a consistency reward is added; and across models, R1 and QwQ produce illegible stretches their non-RL base models do not. Since the base models write legibly and the only systematic intervention is outcome RL, outcome RL can make some models' chains of thought illegible.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 2 supported1 not separately assessed
If the premises hold, the existential conclusion follows with little room for doubt. The weakest point is attributing the effect to outcome RL rather than to some other training or decoding artifact. The within-training ablation, that language consistency declines without a consistency reward, comes closest to isolating RL as the cause. The R1-Zero case, poor readability after pure-RL training, and the cross-model comparison with base models add corroboration. All three premises are still unassessed, though each is drawn from published reports.
Provenance
Where this claim has been said, linked to its canonical form.
We study CoT legibility across 14 reasoning models, finding that RL often causes reasoning to become illegible to both humans and AI monitors, with reasoning models (except Claude) generating illegible CoTs while returning to perfectly readable final answers.
Empirical study of CoT legibility across 14 reasoning models, motivated by CoT monitoring for AI safety.
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.