Training an AI to behave well doesn't teach it new facts — it mostly just unlocks what was always in there.

This reads the question as being about the division of labor between where a model *stores* what it knows and where its *reasoning* operates — and whether alignment-aware training deposits knowledge into the accessible spot, or does something else entirely. The corpus points to a surprising answer: alignment mostly doesn't deposit knowledge at all. It activates knowledge that pretraining already laid down. LIMA is the clean demonstration — just 1,000 carefully curated examples match models trained on orders of magnitude more data, because post-training 'activates existing capabilities rather than building new ones' Can careful curation replace massive alignment datasets?. So the real question isn't whether alignment can *add* knowledge where reasoning can reach it, but whether alignment can shape reasoning without *damaging* the knowledge that's already there.

And here's where the method of alignment matters enormously. Proxy-tuning splits the model apart along exactly the seam the question implies: direct fine-tuning corrupts knowledge storage in the lower layers, whereas decoding-time tuning applies shifts that 'primarily affect reasoning and style' while leaving the base weights — and the knowledge in them — untouched Can decoding-time tuning preserve knowledge better than weight fine-tuning?. That's the mechanism behind the whole question made concrete: knowledge lives in the lower layers, the reasoning/style surface is a distributional shift on top, and clumsy alignment can overwrite the former while trying to adjust the latter. Depositing knowledge 'where reasoning can access it' turns out to be less about placing knowledge and more about not blowing up the deposit that pretraining already made.

The same lesson shows up on the reinforcement-learning side. RLVR doesn't rewrite the model's knowledge — it nudges a tiny minority of high-entropy 'forking' tokens, the pivotal reasoning-decision points, and training on just ~20% of tokens matches full updates Do high-entropy tokens drive reasoning model improvements?. Alignment, in other words, works on the reasoning junctions, not the knowledge substrate. Even stranger: reasoning traces can be deliberately corrupted and still teach as well as correct ones, because they function as 'computational scaffolding rather than meaningful reasoning steps' Do reasoning traces need to be semantically correct?. That reinforces the split — the trace isn't carrying knowledge, it's carrying a procedure for reaching into knowledge that's stored elsewhere.

There's a catch worth knowing you wanted to know. Even when knowledge is safely deposited and reasoning is well-shaped, the reasoning surface is fragile in ways that limit access. Chain-of-thought turns out to be constrained imitation of reasoning *form* — it reproduces familiar patterns and degrades predictably outside the training distribution rather than genuinely reasoning over what the model knows Does chain-of-thought reasoning reveal genuine inference or pattern matching?, Does chain-of-thought reasoning actually generalize beyond training data?. And reasoning accuracy collapses from 92% to 68% with just 3,000 tokens of padding, far below the context limit reasoning-performance-degrades-with-input-length-even-far-below-context-limits. So even perfectly deposited knowledge can be unreachable if the reasoning pathway is imitative or overloaded.

The synthesis, then: alignment-aware training is best understood not as *depositing* knowledge but as tuning the interface to it. The best methods (proxy-tuning, minimal high-quality SFT, entropy-targeted RL) protect the pretrained store and sculpt only the reasoning/style layer that queries it — and where alignment fails, it's usually because it corrupted the store or built a brittle, imitative access path rather than because the knowledge wasn't there.