Extend TAS to larger unknown forget-set cardinalities

Extend Targeted Active Search (TAS) to reliably identify and recover more than two simultaneously forgotten entities when the forget-set cardinality is unknown, addressing the resulting cardinality-estimation challenge.

Background

The experiments primarily evaluate single-entity forgetting and a limited two-entity setting. In the multi-entity experiments, TAS generally ranks forgotten entities above distractors, but its largest-gap estimator sometimes undercounts or overcounts the number of forgotten entities, especially on the denser TOFU benchmark.

The paper therefore identifies scaling to larger and unknown forget-set cardinalities as unresolved. The challenge is not primarily ranking candidate entities—mean average precision is reported as perfect in the tested configurations—but determining where the ranked list should be terminated when several entities may have been unlearned.