Poster: A Preliminary Study of LLM Distillation Inference (opens in a new tab)
Why readLearn how shadow models and hypothesis testing can detect unauthorized LLM distillation with high accuracy.
Researchers proposed a statistical framework to determine if a target LLM was distilled from a proprietary teacher model rather than trained independently. By training paired shadow models on reasoning traces versus reference answers, the auditor converts prediction alignment into a calibrated p-value. In tests using Qwen2.5-7B as teacher and Llama-3.2-3B as suspect, the approach achieved a 1.0 true positive rate at a 0.02 significance level.