Is 4-Bit Quantization More Than Noise to a Chat Model's Safety Behaviour? Attributing Behavioural Shifts with Matched Perturbations
Organizations: University of British Columbia
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 llama.cpp'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.
Figures & tables
| Format | bits/w | NLL/token vs bf16 | ||
|---|---|---|---|---|
| Qwen2.5 | Llama | Qwen3 | ||
| 8bit-g64 | 8.50 | -0.000 | +0.000 | +0.002 |
| 4bit-g32 | 5.00 | +0.108 | +0.026 | +0.104 |
| mixed_4_6 | 4.77 | +0.114 | +0.028 | +0.118 |
| 4bit-g64 | 4.50 | +0.144 | +0.038 | +0.138 |
| MXFP4 | 4.25 | +0.265 | +0.069 | +0.221 |
| 4-bit shift (pp) | Forced start | Free start (held out) |
|---|---|---|
| Qwen2.5-3B | +6.2 | -2.3 |
| Llama-3.2-3B | +0.0 | -10.3 |
| Qwen3-4B | -1.6 | +3.8 |
| XSTest safe | WJB benign | ||
|---|---|---|---|
| Qwen2.5-3B | bf16 | 26.0% | 8.6% |
| 4-bit | -6.4 | -1.0 | |
| Llama-3.2-3B | bf16 | 17.2% | 9.5% |
| 4-bit | -0.4 | +5.7 | |
| Qwen3-4B | bf16 | 17.6% | 21.0% |
| 4-bit | +1.2 | -8.6 |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| Date (2026) | Registered |
|---|---|
| Nov 10 | forced-start primary, depth, judge, churn, format ladder, seed floor |
| Nov 11 | pooled test, two-phase human labels |
| Nov 13 | free-start test on held-out prompts; matched-perturbation controls |
| Nov 14 | other quantizers, dose, refusal direction, GGUF; six random draws; same-prompt bridge |
| D (judge splits) | P (both harmful) | N (neither) | ||||
|---|---|---|---|---|---|---|
| Model | pop. | labelled | pop. | labelled | pop. | labelled |
| Qwen2.5-3B | 101 | 101 | 68 | 25 | 393 | 25 |
| Llama-3.2-3B | 104 | 104 | 120 | 25 | 338 | 25 |
| Qwen3-4B | 45 | 45 | 32 | 25 | 485 | 25 |
| Judge | Llama Guard | Human | |
|---|---|---|---|
| Qwen2.5-3B | +6.2 | +11.0 | +19.4 |
| Llama-3.2-3B | +0.0 | +1.8 | +0.7 |
| Qwen3-4B | -1.6 | -1.4 | -4.0 |
| Model | Cell | bf16 % | build % | Shift [95% CI] | Disc. | Holm | |
|---|---|---|---|---|---|---|---|
| Qwen2.5-3B | 4-bit, | 1.6 | 4.3 | +2.7 [+1.2, +4.3] | 2/17 | ||
| Qwen2.5-3B | 4-bit, | 10.9 | 17.6 | +6.8 [+3.7, +9.8] | 19/57 | ||
| Qwen2.5-3B | MXFP4, | 1.6 | 2.5 | +0.9 [ 0.4, +2.3] | 5/10 | ||
| Qwen2.5-3B | MXFP4, | 10.9 | 14.4 | +3.6 [+0.7, +6.4] | 23/43 | ||
| Qwen2.5-3B | MXFP4, | 18.0 | 17.3 | 0.7 [ 4.1, +2.7] | 49/45 | ||
| Llama-3.2-3B | 4-bit, | 6.4 | 4.4 | 2.0 [ 3.6, 0.5] | 15/4 |
| Model | Cell | bf16 % | build % | Shift [95% CI] | Disc. | Holm | |
|---|---|---|---|---|---|---|---|
| Qwen2.5-3B | 4-bit, XSTest safe | 26.0 | 19.6 | 6.4 [ 11.2, 1.6] | 27/11 | ||
| Qwen2.5-3B | 4-bit, WJB benign | 8.6 | 7.6 | 1.0 [ 4.8, +2.9] | 10/8 | – | |
| Qwen2.5-3B | MXFP4, XSTest safe | 26.0 | 32.0 | +6.0 [+0.8, +11.2] | 16/31 | ||
| Qwen2.5-3B | MXFP4, WJB benign | 8.6 | 10.5 | +1.9 [ 2.4, +6.2] | 8/12 | – | |
| Llama-3.2-3B | 4-bit, XSTest safe | 17.2 | 16.8 | 0.4 [ 4.8, +3.6] | 15/14 | ||
| Llama-3.2-3B | 4-bit, WJB benign | 9.5 | 15.2 | +5.7 [+1.4, +10.0] | 5/17 | – |
| Model | Build | Overall shift [95% cluster CI] | Rewrites base [95% cluster CI] |
|---|---|---|---|
| Qwen2.5-3B | 4-bit | +3.2 [+0.7, +5.7] | +2.9 [ 2.1, +8.1] |
| Qwen2.5-3B | MXFP4 | +3.2 [+0.3, +6.2] | 2.3 [ 9.6, +4.6] |
| Llama-3.2-3B | 4-bit | 5.0 [ 7.8, 2.2] | +1.0 [ 5.6, +7.9] |
| Llama-3.2-3B | MXFP4 | 8.5 [ 11.8, 5.5] | +5.0 [ 1.5, +11.9] |
| Qwen3-4B | 4-bit | +3.8 [+1.8, +6.0] | 1.5 [ 6.2, +2.7] |
| Qwen3-4B | MXFP4 | 1.0 [ 3.3, +1.3] | 2.3 [ 6.2, +1.7] |
| Cell | Build | NLL/token | Norm ratio | Shift [95% CI] |
|---|---|---|---|---|
| Qwen2.5 forced | real 4-bit | +0.144 | – | +6.2 [+2.8, +9.8] |
| dense | +0.144 | 1.000 | +4.8 [+1.2, +8.4] | |
| mirror | +0.131 | 1.000 | 3.0 [ 6.0, +0.0] | |
| permuted 1 | +0.166 | 1.000 | +2.7 [ 0.9, +6.2] | |
| permuted 2 | +0.180 | 1.000 | 1.8 [ 4.8, +1.2] | |
| permuted 3 | +0.155 | 1.000 | +0.9 [ 2.5, +4.3] |
| median relative error | median cosine with RTN error | |||||
|---|---|---|---|---|---|---|
| Model | RTN | MXFP4 | GPTQ | MXFP4 | GPTQ | AWQ |
| Qwen2.5-3B | 0.094 | 0.118 | 0.115 | 0.11 | 0.31 | 0.04 |
| Llama-3.2-3B | 0.093 | 0.118 | 0.111 | 0.07 | 0.32 | 0.13 |
| Model | Build | bf16 % | build % | Shift [95% CI] | Disc. | Holm | |
|---|---|---|---|---|---|---|---|
| Qwen2.5-3B | 4-bit | 48.5 | 53.5 | +5.0 [ 1.5, +11.5] | 19/29 | – | |
| Qwen2.5-3B | MXFP4 | 48.5 | 49.5 | +1.0 [ 6.0, +8.0] | 25/27 | – | |
| Llama-3.2-3B | 4-bit | 45.0 | 30.5 | 14.5 [ 20.5, 8.5] | 34/5 | – | |
| Llama-3.2-3B | MXFP4 | 45.0 | 34.0 | 11.0 [ 17.0, 5.0] | 31/9 | – | |
| Qwen3-4B | 4-bit | 23.5 | 33.5 | +10.0 [+5.0, +15.0] | 4/24 | – | |
| Qwen3-4B | MXFP4 | 23.5 | 23.0 | 0.5 [ 5.0, +4.0] | 11/10 | – |
| Model | bf16 | 8-bit | 4-bit g32 | mixed 4/6 | 4-bit | MXFP4 | MXFP4 deq. |
|---|---|---|---|---|---|---|---|
| Qwen2.5-3B | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 10.5 | 8.5 |
| Llama-3.2-3B | 0.0 | – | – | – | 0.0 | 0.0 | – |
| Qwen3-4B | 0.0 | – | – | – | 0.0 | 0.0 | – |
| Qwen2.5-0.5B | 0.0 | – | – | – | 0.0 | 0.0 | 0.0 |
LLM Quantization
Large language model quantization reduces numerical precision to lower memory use and accelerate training or inference.
See all 443 papers