How sure is the activation oracle?
The paper investigates the confidence and calibration of activation oracles used for interpreting language model outputs. It evaluates six methods for estimating confidence and finds that bootstrap mode frequency is the best-calibrated method. The study highlights the potential of using log-probability as a cost-effective triage signal.
- ▪Activation oracles aim to enhance the interpretability of language model outputs.
- ▪The study tests six methods for estimating the confidence of activation oracles on 6,000 samples.
- ▪Bootstrap mode frequency outperforms other methods in terms of calibration.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computation and Language arXiv:2605.26045 (cs) [Submitted on 25 May 2026] Title:Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals Authors:Federico Torrielli, Peter Schneider-Kamp, Lukas Galke Poech View a PDF of the paper titled Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals, by Federico Torrielli and 2 other authors View PDF HTML (experimental) Abstract:Activation oracles aim to make the activations of other models legible to humans and yield promising results compared to white-box interpretability techniques. However, uncertainty quantification (UQ) for the natural-language outputs of such activation oracles is so far understudied.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
