Preprint finds AI memory compression can weaken safety refusals

Researchers report in an arXiv preprint that compressing the memory used during language model inference can weaken safety refusals while leaving standard quality measures largely unchanged. Across eleven instruction-tuned models and five benchmarks containing 1,894 prompts, Mistral-7B lost 15.2% of its refusals at only 1.03x perplexity, a measure of prediction error. The researchers propose a diagnostic that identifies how compression damages safety-related features and guides mitigation. Their training-free protocol requires approximately 35 GPU-minutes and recovers up to 97% of lost alignment with under 1% latency overhead and minimal memory overhead. The study finds no universal safe bit-width: failure thresholds differ by model.