Can this principle apply to other intermediate text generation tasks?
This explores whether the idea that 'intermediate' generated text (reasoning steps, scratchpads, chains of thought) can be decoupled from — or doesn't actually require — visible token-by-token verbalization extends to other tasks where text is produced as a means rather than an end.
This reads the question as: if reasoning tokens turn out to be more of a training artifact than a true requirement, does that lesson generalize to other places where models generate text as an intermediate step rather than a final product? The corpus has a surprisingly coherent answer, and it cuts both ways. Several notes converge on the claim that the visible intermediate text is often not where the real work happens. Can models reason without generating visible thinking tokens? shows depth-recurrent and latent architectures (Coconut, Heima) scaling reasoning entirely in hidden state, suggesting verbalization is optional. Do transformers hide reasoning before producing filler tokens? goes further: models can compute the correct answer in early layers and then actively overwrite it with format-compliant filler — the intermediate text is theater, recoverable only from lower-ranked predictions. If the principle is 'the useful computation can be separated from the emitted intermediate string,' these say yes, strongly.
The diffusion-model angle generalizes the principle most cleanly. Can reasoning and answers be generated separately in language models? shows that once you drop the left-to-right constraint, reasoning and answer become two refinement axes rather than a sequence — reasoning can keep refining after the answer has already converged. That reframes 'intermediate text' as a knob (how much refinement compute to spend) rather than a fixed prefix you must emit before the answer. Any intermediate-generation task — planning, query rewriting, draft-then-edit — could in principle be restructured this way.
But the same corpus warns where the principle breaks. Does AI text generation unfold through temporal reflection? and Does LLM generation explore competing claims while producing text? both argue that generated text, intermediate or not, is a smooth probabilistic continuation, not genuine deliberation — it doesn't explore counterpositions or spend reflective time. So if your 'intermediate task' was secretly relying on the visible text to do real exploratory work, removing or hiding it doesn't recover what was never there. The text wasn't reasoning; compressing it just makes that explicit.
There's a harder limit worth knowing about. The principle assumes the intermediate step is reorganizing capability the model already has. Can prompt optimization teach models knowledge they lack? and What stops large language models from improving themselves? mark the ceiling: intermediate generation — whether verbalized, latent, or refined in place — can only activate and route existing knowledge, not supply what's missing or self-validate beyond the generation-verification gap. So the principle transfers to other intermediate text tasks precisely when those tasks are about activation and routing (reasoning traces, retrieval planning, draft scaffolding) and fails when the intermediate text was being asked to do something it never could — generate new knowledge or verify itself.
The quietly useful takeaway: 'intermediate text' is doing one of two very different jobs — either it's a load-bearing compute scaffold (in which case it can often be moved into latent space or refined out of the prefix) or it's a contract with the outside world (an auditable trace, a verifiable write-back as in Can RAG systems safely learn from their own generated answers?). The principle generalizes freely for the first job and not at all for the second — and most of the disagreement about whether 'hidden reasoning' is safe is really a disagreement about which job the text was doing.
Sources 8 notes
Multiple architectures—depth-recurrent models, Heima, and Coconut—demonstrate that test-time compute scales through hidden state iteration rather than token generation. This suggests verbalization is a training artifact, not a reasoning requirement.
Logit lens analysis shows models trained with hidden CoT tokens compute correct answers in layers 1-3, then actively suppress these representations in final layers to produce format-compliant filler output. The reasoning is fully recoverable from lower-ranked token predictions.
ICE shows that bidirectional attention in diffusion LLMs enables in-place prompting—embedding reasoning directly in masked positions refined alongside answers. Answer confidence converges early while reasoning continues refining, allowing early-exit mechanisms to cut compute by 50% while maintaining accuracy.
Token ordering in LLMs follows probabilistic selection without intervening reflection or revision. Human discourse gains meaning from temporal structure—time spent thinking changes what comes next—but AI text production lacks this duration-in-reflection despite appearing sequentially composed.
Token prediction trains models to continue toward the training distribution, not to explore logically related counterpositions. This smoothness in process produces smooth claims that multiply without generating new perspectives.
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
Self-improvement in LLMs is formally bounded by the generation-verification gap, meaning every reliable fix requires something external to validate and enforce it. Models cannot escape this constraint through metacognition alone.
Systems can add generated answers to their retrieval corpus when outputs pass entailment verification, source attribution checks, and novelty detection. This prevents hallucinations from polluting future retrievals while allowing genuine knowledge accumulation.