argumentA line of reasoning from premises to a conclusion. Like a claim, it belongs to no one who makes it.constitution →

The argument from held-out-fact recovery experiments

forThis argument, if it holds, bears in favour of the claim.constitution →Current LLM unlearning methods leave unlearned pretraining knowledge largely recoverable by fine-tuning on other facts from the forget set.

The argument

When a model unlearned on a set of mutually independent pretrained facts is fine-tuned on part of that set, its accuracy on the held-out part returns to at least 88% of what the original model reaches under the same fine-tuning, for RMU, gradient difference and random-incorrect-answer training. The fine-tuning data cannot teach the held-out facts, so they were still encoded in the weights. Other attacks show the same fragility across standard methods: brief fine-tuning restores knowledge these methods were applied to remove. The methods tested are representative of those in current use, so current methods leave such knowledge largely recoverable.

Does it work?

Does it follow?: likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
Do the premises hold?: 1 supported

Granting its premises, the conclusion follows for the methods tested and is made much more likely, though not guaranteed, for standard methods generally. The step from held-out recovery to "still in the weights" depends on the forget facts being independent, which the experiment controls by construction for synthetic facts and by curation and a matched baseline for pretrained ones. The weight falls on representativeness: the direct result covers three methods on small models, and its extension rests on independent findings that brief fine-tuning reverses standard unlearning, which are consistent across several groups but use different attacks. The argument does not reach methods designed to rebuild the model without the knowledge.

Premises

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →Brief fine-tuning restores knowledge that standard LLM unlearning methods such as RMU and gradient ascent were applied to remove.

Where it is made

Passages where sources make, criticize or mention this argument. They record where it appears, not how strong it is, and none has precedence over another.

Do Unlearning Methods Remove Information from Language Model Weights?makes itstates it fullyThe passage states the inference with what it needs.constitution →
we give an attacker access to some facts that were supposed to be removed, and using those, the attacker tries to recover other facts from the same distribution that cannot be guessed from the accessible facts. We show that using fine-tuning on the accessible facts can recover 88% of the pre-unlearning accuracy when applied to current unlearning methods for information learned during pretraining, revealing the limitations of these methods in removing information from the model weights.