A small language model detects behavioural faithfulness gaps that frontier judges and human raters miss

Abstract:Whether a language model behaves as it claims is a judgement on which independent human raters cannot agree (Fleiss kappa = 0.074). We show that a small, purpose-built instrument does better. A linear read-out of the frozen representation of a from-scratch 146-million-parameter auditor detects companion-induced sycophancy, dependence-fostering and confabulated memories on generator models unseen in training (AUROC 0.804, leave-one-generator-out, against ground truth fixed at generation, independent of human judgement), where a frontier zero-shot judge on the identical items reaches 0.721 and falls to chance on the most distant family. The auditor's substrate is hyperbolic, and its demonstrated benefit is hierarchical: an ablation isolates the advantage over a matched Euclidean control on multi-domain structure. On this task, behavioural faithfulness is measured not by scale but by a small, purpose-built instrument.

Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC); Machine Learning (cs.LG)

arXiv:2607.09306 [cs.CL]

(or arXiv:2607.09306v2 [cs.CL] for this version)

https://doi.org/10.48550/arXiv.2607.09306 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.