Weird Generalization is Weirdly Brittle

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wanner, Miriam, Collison, Hannah, Jurayj, William, Van Durme, Benjamin, Dredze, Mark, Walden, William
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913086698422272
author Wanner, Miriam
Collison, Hannah
Jurayj, William
Van Durme, Benjamin
Dredze, Mark
Walden, William
author_facet Wanner, Miriam
Collison, Hannah
Jurayj, William
Van Durme, Benjamin
Dredze, Mark
Walden, William
contents Weird generalization is a phenomenon in which models fine-tuned on data from a narrow domain (e.g. insecure code) develop surprising traits that manifest even outside that domain (e.g. broad misalignment)-a phenomenon that prior work has highlighted as a critical safety concern. Here, we present an extended replication study of key weird generalization results across an expanded suite of models and datasets. We confirm that surprising (and dangerous) traits can emerge under certain circumstances, but we find that weird generalization is exceptionally brittle: it emerges only for specific models on specific datasets, and it vanishes under simple training-time, prompt-based interventions. We find that the most effective interventions provide prompt context that makes the generalized behavior the expected behavior. However, we show that even very generic interventions that do not anticipate specific generalized traits can still be effective in mitigating weird generalization's effects. Our findings thus help clarify the nature of the safety threat that weird generalization poses and point toward an easily implemented set of solutions.
format Preprint
id arxiv_https___arxiv_org_abs_2604_10022
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Weird Generalization is Weirdly Brittle
Wanner, Miriam
Collison, Hannah
Jurayj, William
Van Durme, Benjamin
Dredze, Mark
Walden, William
Computation and Language
Weird generalization is a phenomenon in which models fine-tuned on data from a narrow domain (e.g. insecure code) develop surprising traits that manifest even outside that domain (e.g. broad misalignment)-a phenomenon that prior work has highlighted as a critical safety concern. Here, we present an extended replication study of key weird generalization results across an expanded suite of models and datasets. We confirm that surprising (and dangerous) traits can emerge under certain circumstances, but we find that weird generalization is exceptionally brittle: it emerges only for specific models on specific datasets, and it vanishes under simple training-time, prompt-based interventions. We find that the most effective interventions provide prompt context that makes the generalized behavior the expected behavior. However, we show that even very generic interventions that do not anticipate specific generalized traits can still be effective in mitigating weird generalization's effects. Our findings thus help clarify the nature of the safety threat that weird generalization poses and point toward an easily implemented set of solutions.
title Weird Generalization is Weirdly Brittle
topic Computation and Language
url https://arxiv.org/abs/2604.10022