Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals

Federico Torrielli, Peter Schneider-Kamp, Lukas Galke Poech · 2026

Read Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals by Federico Torrielli, Peter Schneider-Kamp, Lukas Galke Poech (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-26045.txt.

Description

Torrielli, Schneider-Kamp and Galke Poech test trust in activation oracles. Such an oracle turns another model's inner activations into words, with no confidence score. The authors compare five confidence methods across four Qwen and Gemma oracles, sized eight to twenty-seven billion parameters. The test is a secret-word task: recover a word trained into a model but missing from its outputs. With six thousand samples per pairing, the best method hinges on whether candidate answers can be listed, or a calibration set is at hand.

Catalog record

CREATORFederico Torrielli, Peter Schneider-Kamp, Lukas Galke Poech
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59674
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.26045

Browse the full library · About the Tower of Babel