Compressing an AI model to run cheaper does not automatically make it less safe — but cranking up its randomness might.
Researchers ran a factorial experiment across 9 instruction-tuned models, three precision levels (FP16, INT8, INT4), and six sampling temperatures, generating roughly 322,000 responses scored by a six-model safety ensemble. The headline finding: standard INT4 quantization held or improved safety for 7 of the 9 models tested. The exception was SmolLM3-3B, a weaker baseline whose attack success rate nearly doubled, from 18.5% to 36.0%. That outlier aside, low-bit compression alone was not the villain the field often assumes it to be.
The more interesting threat is temperature. At T=1.0 — a setting many deployments use to avoid repetitive outputs — decision instability hit 53.0% for vulnerable models even when average attack success rates looked acceptable. That gap matters: a safety evaluation that reports only average pass rates can look fine while the model is, in practice, unsafe roughly half the time. The paper calls this metric Decision Flip Rate, and it deserves wider adoption in standard benchmarks.
Critically, the two knobs do not compound each other in the way a pessimist might expect. The Compound Degradation Index stayed largely sub-additive, meaning the combined effect of quantization plus elevated temperature is smaller than the sum of its parts. The practical upshot: teams running strongly aligned models on INT4 can stop worrying about compression as a primary risk vector — but anyone reporting safety results at high temperature owes readers multi-sample stability numbers, not just averages.