What question did this study set out to answer?

This analysis aims to evaluate the effectiveness of honesty probes against knowledge-preserving weight attacks in language models.

July 10, 2026Open Access

Fathom v28 / styxx v7.24.3: Calibration Poisoning, Not Erasure (v0.3) -- Substrate Honesty Probes Survive Knowledge-Preserving Weight Attacks When the Auditor's Calibration Is Private

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial examines honesty probes against knowledge-preserving attacks, indicating robust defense mechanisms.

This analysis aims to evaluate the effectiveness of honesty probes against knowledge-preserving weight attacks in language models.
Utilized linear probes on a language model's residual stream to assess model honesty.
Employed a knowledge-preserving LoRA attacker and analyzed the performance of the probe under various conditions.
Verified the effectiveness of the private clean calibration against adaptive attackers.
The naive static attacker displayed an AUROC of 0.711-0.838, illustrating a significant protective capability.
Compared to adaptive schemes, the naive attacker proved to be the strongest threat identified, with a higher calibrated read of 0.833.
Calibration poisoning, rather than erasure, was confirmed, ensuring the integrity of the honesty probe readings.

Alexander Rodabaugh (2026) studied this question.

View Full Paper

Ask AI

Bookmark