Search papers, labs, and topics across Lattice.
Affiliation:
2
0
4
4
Probe of Internal Recognition (PIR) separates a model that will not answer from one that cannot, which supports sandbagging audits and unlearning verification, and extends from multiple-choice questions to free-form generation.
High reconstruction scores in activation explanations don't guarantee truthfulness, but RECAP ensures that designated content remains verifiable and decodable, even under adversarial manipulation.
