- Article
23 Pages
Hidden-state probes are widely used to study safety-related behavior in language and multimodal models, yet their results can change with the data sources, behavior labels, extracted layer, and probe family. We introduce TriSAIL, an evaluation protocol that records these choices and tests how they affect the resulting claim. TriSAIL uses three response-derived states—benign, non-refusal jailbreak, and borderline/refusal—and assigns train, validation, and test data by source. We extract hidden states from the final input position in the first generation forward pass, before a generated token is returned. The operating layer is selected using validation data only. We evaluate the protocol on five text-only LLMs and six MLLMs with logistic, kNN, and SVM probes. The LLM results vary substantially across held-out attack families, probe families, and source assignments. The MLLM probes achieve high label separability, while source-ID decoding, prediction–source association, and residualization tests show strong source alignment in the same representations. Because the main MLLM matrix couples sources and labels, these results describe source-sensitive separability. Source-transfer conclusions use source-held-out columns; mixed accuracy is a descriptive overall summary. TriSAIL reports these results with probe sensitivity and source diagnostics so that each score is interpreted under the conditions that produced it.
Symmetry
30 September 2026






![Energy of positive- and negative-parity states in the
Ca
40
nucleus as a function of J. The states identified with small green circles correspond to well-established spin–parity assignments, while the small red circles indicate states with uncertain spin–parity assignments. The brown lines connecting the circles represent electromagnetic transitions. All the states shown in the figure are taken from the compilation [59,60]. The black lines indicate rotational bands. In (a), the bands built on the 0+, 8+, and 3+ states, as well as the SD band, were previously proposed [9,59,60]. In (b), the 0− band was previously suggested [11,59,60]. The arranged bands are labeled according to the spin–parity (
J
π
) of their band head, with subscripts used to distinguish between bands with the same bandhead spin–parity.](https://mdpi-res.com/cdn-cgi/image/width=281%2Cheight=192/https://mdpi-res.com/symmetry/symmetry-18-01193/article_deploy/html/images/symmetry-18-01193-g001a-550.jpg)




