What question did this study set out to answer?

To analyze how language models can report false information while retaining accurate answers through various corruption channels.

August 1, 2026Open Access

Fathom v31 / styxx: Frame-Locality -- Where Corruption Captures a Language Model's Report, and Where It Reaches the Belief

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial evaluates the effects of corruption on language model reporting, suggesting significant implications for AI frameworks.

To analyze how language models can report false information while retaining accurate answers through various corruption channels.
Performance evaluated against multiple inference-time channels
Weight-channel manipulation tested using distinct models
Recalls measured using OATH certification for reproducibility
Out-of-frame recovery at 0.975 for uncorrupted responses, and 0.696 for corrupted ones, indicating a significant impact of corruption.
Knowledge-preserving attacks showed a recovery rate of about 0.5111, preserving most of the underlying belief without measurable loss to capability.
Demonstrated that certain attacks can overwrite model beliefs, leading to reduced general capability metrics.

Alexander Rodabaugh (2026) studied this question.

View Full Paper

Ask AI

Bookmark