Useful Features, Backward Scores: OOD in Language-Model Trajectories

Abstract:Out-of-distribution (OOD) detectors prioritize inputs for closer inspection. Yet features that distinguish input groups need not yield a useful anomaly ranking. We analyze this gap in language-model trajectories under text-length control and fixed score directions. On Spam development data, an input adaptation of D^2HScore falls from raw AUROC 0.919 to 0.530 after length matching. On length-matched, held-out HateSpeech inputs, the same features yield AUROC 0.644 for a labeled linear classifier but 0.444 for an ID-fitted distance score. ToxicChat shows the same contrast. Feature-selection and backbone controls retain the main reversal pattern. Frozen Civil Comments and TweetEval irony tests also reverse (0.467 and 0.435), extending the finding beyond toxicity. In these contrasts, anomalous groups have farther centers but tighter spread. A labeled, fixed-center feature-space intervention changes rankings: equalizing spread helps some tasks and harms others. OOD evaluation must check the chosen score's ranking even when its features distinguish the classes.

Computation and Language (cs.CL); Machine Learning (cs.LG)

arXiv:2605.00269 [cs.CL]

(or arXiv:2605.00269v2 [cs.CL] for this version)

https://doi.org/10.48550/arXiv.2605.00269 Focus to learn more arXiv-issued DOI via DataCite

code_supplement/README.md

code_supplement/manifest.json

code_supplement/numeric/CivilTest.npy

code_supplement/numeric/CivilTest_labels.npy

code_supplement/numeric/CivilTest_lengths.npy

code_supplement/numeric/CivilTest_scores.npy

code_supplement/numeric/CivilValidation.npy

code_supplement/numeric/HateSpeech.npy

code_supplement/numeric/HateSpeech_bootstrap_aurocs.npy

code_supplement/numeric/HateSpeech_distances.npy

code_supplement/numeric/HateSpeech_labels.npy

code_supplement/numeric/HateSpeech_lengths.npy

code_supplement/numeric/HateSpeech_scores.npy

code_supplement/numeric/Jailbreak.npy

code_supplement/numeric/Jailbreak_bootstrap_aurocs.npy

code_supplement/numeric/Jailbreak_distances.npy

code_supplement/numeric/Jailbreak_labels.npy

code_supplement/numeric/Jailbreak_lengths.npy

code_supplement/numeric/Jailbreak_scores.npy

code_supplement/numeric/Spam.npy

code_supplement/numeric/Spam_bootstrap_aurocs.npy

code_supplement/numeric/Spam_distances.npy

code_supplement/numeric/Spam_labels.npy

code_supplement/numeric/Spam_lengths.npy

code_supplement/numeric/Spam_scores.npy

code_supplement/numeric/ToxicChat.npy

code_supplement/numeric/ToxicChat_bootstrap_aurocs.npy

code_supplement/numeric/ToxicChat_distances.npy

code_supplement/numeric/ToxicChat_labels.npy

code_supplement/numeric/ToxicChat_lengths.npy

code_supplement/numeric/ToxicChat_scores.npy

code_supplement/numeric/civil_bootstrap_estimates.npy

code_supplement/numeric/features.npy

code_supplement/numeric/id.npy

code_supplement/numeric/labels.npy

code_supplement/numeric/lengths.npy

code_supplement/numeric/ood.npy

code_supplement/numeric/reference.npy

code_supplement/numeric/roles.npy

code_supplement/numeric/values.npy

code_supplement/paper/verify_code_supplement.py

code_supplement/src/utils/radial_analysis.py

(37 additional files not shown) You must enabled JavaScript to view entire file list.

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.