Does Abliteration Change the Effect of Cybersecurity Fine-Tuning?

Under peer review. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Does making a language model less likely to refuse a request help it learn from cybersecurity training? We compared two versions of Qwen 3.5 4B, one original and one edited to reduce refusals, before and after the same training. On 850 cybersecurity multiple-choice questions, the original model returned 10 more correct answers after training. The edited model returned 312 fewer. The scoring rule required an answer in the requested format. Even with 16 times as much room to respond, the trained, edited model returned 614 answers without a valid choice. The study leaves open how many of those answers a human reader could interpret correctly. For practitioners evaluating a model for a security workflow, the distinction matters. A model can be willing to answer and still fail to give a tool the answer it needs. Testing the starting model and the trained one on the same questions makes that change visible. The study covers one model family and one training run per starting state.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24358967
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Does Abliteration Change the Effect of Cybersecurity Fine-Tuning?

Alexander Giannini, Eric Allione
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
preprint

Does Abliteration Change the Effect of Cybersecurity Fine-Tuning?

Alexander Giannini, Eric Allione
preprint en

Abstract

Under peer review. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Does making a language model less likely to refuse a request help it learn from cybersecurity training? We compared two versions of Qwen 3.5 4B, one original and one edited to reduce refusals, before and after the same training. On 850 cybersecurity multiple-choice questions, the original model returned 10 more correct answers after training. The edited model returned 312 fewer. The scoring rule required an answer in the requested format. Even with 16 times as much room to respond, the trained, edited model returned 614 answers without a valid choice. The study leaves open how many of those answers a human reader could interpret correctly. For practitioners evaluating a model for a security workflow, the distinction matters. A model can be willing to answer and still fail to give a tool the answer it needs. Testing the starting model and the trained one on the same questions makes that change visible. The study covers one model family and one training run per starting state.

Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon