Why do fine-tuned models fail to use facts they memorize?

Fine-tuning to inject new knowledge produces a puzzling split the authors name the Knowing–Using Gap: a model reliably recalls a freshly memorized fact ("Sydney is located in [Entity]") yet fails the multi-hop question that depends on it ("the capital of the country Sydney is in is..."). There is even a temporal lag — memorization lands well before generalization does, if it arrives at all. The tempting explanation is a capacity or data-fit problem, but this work rules that out.

The mechanism is a causal intervention called self-patching: it searches for activation locations where relocating a memorized representation substantially rescues a failed generalization case. That such locations exist is the whole argument — the knowledge is present in the network, just encoded in storage-oriented early or late layers, and it never permeates into the mid-layer circuits where reasoning actually runs. Therefore the gap is a routing problem, not a capacity one: the fact resides in the model, just not where computation can reach it. And it is partially reversible — a simple heuristic built on the diagnostic recovers 58–75% of the oracle headroom, replicating across models, scales, and domains.

This mechanistically deepens an existing behavioral finding. Since Why does reasoning training help math but hurt medical tasks?, the layer-localization of knowledge vs. reasoning was already visible; self-patching shows the failure isn't just where knowledge lives but that fine-tuning deposits it in the wrong tier. And since Does procedural knowledge drive reasoning more than factual retrieval?, the routing story explains why injected facts behave like inert retrieval rather than usable procedure — they never reach the circuits that generalize. The reframing is practical: fine-tuning's generalization failure becomes a tractable alignment problem, pointing to alignment-aware training that deposits knowledge where reasoning can use it.