distillation-detection

#distillation-detection

Reference-Based Distillation Detection in LLMs

arXiv cs.LG ↗ · 6d ago Cached

This paper introduces a reference-based method to detect whether an LLM was distilled from a specific teacher model, using membership inference. The approach achieves near-perfect accuracy in controlled settings and provides new evidence about potential distillation relationships involving QwQ, DeepSeek-R1, and GPT-OSS.

0 favorites 0 likes

#distillation-detection

@qqqqqf_: I additionally tested Fable5. I found that it does not achieve very high similarity (>0.9) with most of Anthropic's models. This likely indicates that it could be an unprecedented new base model trained on a mixture of data (just speculation; note that post-training also affects fingerprint results, for reference only). But there is an extra surprise…

X AI KOLs Timeline ↗ · 2026-07-02 Cached

User @qqqqqf_ shared fingerprint test results for Fable5 and other models, finding that Fable5 has low similarity with most Anthropic models but very high similarity with kimi k2.7, qwen3-7max, glm5.2, etc., speculating possible data leakage or distillation. At the same time, they released an open source project ai-model-fingerprint for model identification and detection.

0 favorites 0 likes

distillation-detection

Reference-Based Distillation Detection in LLMs

Submit Feedback