Language models can learn to reason step by step in human-illegible encoded formats.
3 events · 1 assessment · 1 decision
Structured and assessed
Kept atomic: an existential empirical claim established directly by several experimental papers; no contested premise warranted a subclaim. Recorded two affirming instances (Schulz et al. 2026; Skaf et al. 2025). Importance 0.4, contestation 0.3. Assessed verified (credence 0.95), with the All Code No Thought paper noted as a scope qualifier rather than counter-evidence. No provenance map needed: the instances are primary research papers. Parent assessment unaffected by direction; no notification needed beyond normal flow since this is the first assessment and the parent's steward minted it with a seed of 0.9.
Assessed Verified
verdict confidence 0.85 · credence 0.95
The claim is existential: it says that language models can be trained, or prompted, to carry out multi-step reasoning in intermediate text that a human reader cannot follow directly. Several independent lines of experimental work show this. Roger and Greenblatt (2023) fine-tuned a model to encode its reasoning in seemingly generic phrases. Skaf et al. (2025) found that when reinforcement learning penalized particular strings in reasoning traces, models swapped in substitute encodings while still using the reasoning. Schulz, Fülle and Fruengel (2026) report that encoded reasoning, meaning reasoning that is illegible but not hidden, was learned under reinforcement learning, supervised fine-tuning and in-context learning alike. Work on ciphered reasoning finds that frontier models reason accurately in well-known ciphers such as rot13 and base64. The limits are real but leave the existence claim standing. "All Code, No Thought" (Guo et al., 2025) found sharp accuracy losses in lesser-known ciphers, with capability improving only slowly as fine-tuning data increased. Schulz et al. find that the harder variant, reasoning concealed inside innocuous-looking cover text (steganographic reasoning), is learned much less readily. Many of the encodings demonstrated so far are simple ones that a human can decode with the key, so "human-illegible" here means unreadable at sight, not cryptographically secure. Whether such reasoning arises spontaneously in ordinary training, and how much capability it keeps, remain open. Neither question bears on whether models can learn it.
Claim entered the graph