Unlearning audit predicts when removed model knowledge can recover

A new arXiv paper asks whether language-model unlearning audits can do more than check whether a model currently refuses or fails to output target knowledge.

The researchers introduce J-Access, an inference-time audit that uses the Jacobian lens to map intermediate representations into vocabulary space. It measures how often target concepts remain accessible along the model’s output pathway after unlearning.

The key hypothesis is that residual accessibility predicts recovery risk. If target knowledge remains close to the output pathway, less continued training may be needed to restore it. The authors audit 398 public unlearned models across eight unlearning methods and report that many retain access above a retain-only baseline.

The result matters for AI safety because a model can appear to have forgotten information while still carrying internal traces that later fine-tuning can revive. The paper frames internal auditing as a tool for proactive risk monitoring, not just post-hoc evaluation.