Newshot

Is 4-Bit Quantization More Than Noise to a Chat Model's Safety Behaviour? Attributing Behavioural Shifts with Matched Perturbations

Summary

arXiv:2611.22186v1 Announce Type: cross Abstract: Chat models on phones and laptops are usually 4-bit copies of the model that was evaluated, and audits of such copies disagree about what quantization does to refusal. A 4-bit build is one of many weight perturbations of the same size, so a change in its behaviour could come from its specific rounding error or from any perturbation that large. In a pre-registered study of three 3-4B chat models, with paired tests, blind human labels and an 8-bit null, we find two 4-bit shifts of opposite sign: with a compliant reply prefilled, Qwen2.5-3B continues into harmful content more often, and on held-out jailbreaks Llama-3.2-3B complies 10.3 points less often. Matched perturbations separate the two. Shuffled and Gaussian versions of the error average close to zero, while scaling the real error, from its mirror image to twice its size, changes behaviour near-linearly. At half the real error, Llama's likelihood is almost unchanged, yet it already…

Original text(via arXiv cs.CL)

Computer Science > Cryptography and Security

Title:Is 4-Bit Quantization More Than Noise to a Chat Model's Safety Behaviour? Attributing Behavioural Shifts with Matched Perturbations

Abstract:Chat models on phones and laptops are usually 4-bit copies of the model that was evaluated, and audits of such copies disagree about what quantization does to refusal. A 4-bit build is one of many weight perturbations of the same size, so a change in its behaviour could come from its specific rounding error or from any perturbation that large. In a pre-registered study of three 3-4B chat models, with paired tests, blind human labels and an 8-bit null, we find two 4-bit shifts of opposite sign: with a compliant reply prefilled, Qwen2.5-3B continues into harmful content more often, and on held-out jailbreaks Llama-3.2-3B complies 10.3 points less often. Matched perturbations separate the two. Shuffled and Gaussian versions of the error average close to zero, while scaling the real error, from its mirror image to twice its size, changes behaviour near-linearly. At half the real error, Llama's likelihood is almost unchanged, yet it already complies 5.8 points less often, and this http URL's Q4_K_M, at the same bits per weight, moves neither model. Refusal under 4-bit quantization depends on the specific rounding error, not on bits per weight or likelihood, so each build has to be compared with its original.
Comments:16 pages, 4 figures. Code and labels: this https URL
Subjects:Cryptography and Security (cs.CR) ; Computation and Language (cs.CL)
ACM classes:I.2.7
Cite as:arXiv:2611.22186 [cs.CR]
(or arXiv:2611.22186v1 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2611.22186 arXiv-issued DOI via DataCite (pending registration)

Submission history

Access Paper:

  • HTML (experimental)

Additional Features

  • Audio Summary

Current browse context:

References & Citations

  • Google Scholar
  • Semantic Scholar

BibTeX formatted citation

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

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.

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by arXiv cs.CL
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗