Paper: fine-tuning failure is a routing problem, not storage

Fine-tuning is supposed to be the reliable way to give a language model new facts, but there is a stubborn failure mode that shows up in practice. The model regurgitates the injected fact on a direct probe, then flunks anything that requires actually using it. A new arxiv preprint from Lu Dai and colleagues names this the Knowing-Using Gap and argues it is not really a storage problem at all.

Their diagnosis is that memorized representations are landing inside the network but not reaching the layers where reasoning happens. They frame it as a knowledge-circuit misalignment hypothesis: the fact is in there, it just is not routed to a computation-effective layer. To probe this, the team introduces a technique they call self-patching, which relocates activations during a forward pass to see where the useful representation actually needs to be for the model to use it downstream.

The payoff they report is a simple heuristic strategy, derived from the diagnosis, that recovers 58 to 75 percent of what they call the oracle headroom on generalization failure, with experiments run cross-domain. That is a large chunk of the ceiling from a fix that starts as a mechanistic hypothesis, which is why the result matters even if you do not build models yourself. The standard fine-tuning recipe may be leaving obvious performance on the table because the field has been optimizing for the wrong signal, namely whether the model can recite the injected fact rather than whether it can use it.

The honest caveat is that this is a preprint reporting one team's method, the oracle headroom framing is the ceiling the authors themselves define, and cross-domain here means the domains the authors chose. What the paper does not give you is a plug-in fix for a specific open-weight model, a scaling curve to frontier sizes, or a compute overhead number for running the diagnostic inside a real training loop.

If the underlying claim survives replication, the winners are the enterprise teams that have been quietly discounting fine-tuning as a knowledge-injection method because the memorized facts did not survive contact with downstream tasks. A routing-aware curriculum would be considerably cheaper than the alternative, which is spending more compute training the same fact in until it happens to route correctly.