When a language model stops producing the expected answer, the simplest reading is that it has lost the information. An experiment on Qwen2.5-1.5B suggests a different interpretation: the fact may remain encoded in a form that does not immediately translate into behavioral accuracy, and the speed at which the model reacquires it reveals the trace.

The researchers compared three counterbalanced training histories. In LEARNED, the model receives the correct factual binding. In CONTROL, it receives matched exposure without that binding. In WRONG, it receives a competing relation-valid answer in the same slot. After an identical rehearsal-free continual fine-tuning stream, LEARNED reacquires the target faster than CONTROL at a prespecified checkpoint of partial forgetting. WRONG shows no advantage over CONTROL, while the comparison with LEARNED provides slot-matched corroboration, with possible competitor interference. The advantage remains positive after adjustment for measured pre-relearning behavior. In a secondary descriptive conversion using an unpaired never-seen pool, prior learning reduces the optimizer updates required for restoration by about 18%.

The point is not merely academic. For teams managing LLMs locally, continual fine-tuning is often a compromise between updating the model with new facts and avoiding degradation of previous knowledge. Rehearsal-free pipelines are attractive because they avoid retaining past examples, a relevant constraint where data sovereignty and minimization are priorities. But the downside is forgetting. This study indicates that behavioral forgetting is not equivalent to complete erasure of the trace: a relearning saving tied to the specific binding remains. It is not generic familiarity with the entity, vocabulary, or query format, because CONTROL does not benefit from the same advantage. Nor is it the effect of having seen a competing answer, because WRONG does not accelerate recovery.

This distinction changes how updates should be evaluated. A model may look identical to another after a certain number of steps, but if one of them has already encountered the correct binding in the past, the restoration cost may be lower. Current accuracy as the only metric therefore underestimates the lasting effect of previous experience. In TCO terms, this affects the number of optimization runs, machine time, and the amount of data to be relabeled. The 18% saving in required updates, although descriptive and secondary, signals that reuse of latent structure is not negligible.

There is also the opposite side: if an organization must remove information for regulatory or security requirements, getting the model to stop producing it does not guarantee that the trace has been erased from the parameters. The ease with which a later update can restore a previously learned binding is a signal worth watching.

The result suggests separating two questions that are often confused: did the model lose the fact, or did it merely become less accessible? The answer matters for those designing update cycles on their own infrastructure, because it shifts attention from accuracy alone to the marginal cost of restoration.