Poster: A Preliminary Study of LLM Distillation Inference
The study tested if a suspect LLM was distilled from a teacher model by framing it as a hypothesis test. By training shadow models to represent distilled and independently trained behaviors, the method used three signals—log-probability, entropy, and token agreement—to score suspects. Aggregating these scores created a calibrated likelihood-ratio test that successfully detected every distilled suspect model with 100% true positive rate.