Skip to content
NewA calmer, faster ZeroHour

4-bit rounding error, not bit width, shifts chat-model refusal

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

arXiv cs.AI / cs.LG / cs.CLby Ege Kaya ()1 min read1 · 1 read
AI summary · x-ai/grok-4.7Summaries by Vani AI

Qwen2.5-3B continues harmful replies more often, while Llama-3.2-3B complies 10.3 points less on held-out jailbreaks.

A pre-registered study of three 3–4B chat models tests whether 4-bit refusal changes come from a specific rounding error or from any equally large weight perturbation. With a compliant reply prefilled, Qwen2.5-3B continues into harmful content more often; on held-out jailbreaks, Llama-3.2-3B complies 10.3 points less often. Shuffled and Gaussian errors average near zero, while scaling the real error changes behaviour near-linearly. llama.cpp Q4_K_M, at the same bits per weight, moves neither model.

Key facts
  • At half the real error, Llama's likelihood is almost unchanged but compliance falls 5.8 points.
  • An 8-bit null and paired tests with blind human labels were used.
  • The authors say each 4-bit build must be compared with its original.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at arxiv.org

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.

arXiv cs.AI / cs.LG / cs.CL

By Ege Kaya · arxiv.org

Read the full article at arXiv cs.AI / cs.LG / cs.CL

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.