Bounded Behavioral Indistinguishability for Black-Box LLM Distillation

In three linesStudy of black-box LLM distillation through bounded behavioral indistinguishability. Authors evaluate Qwen and Llama pairs with 5,000-prompt suite, showing LoRA improves semantic similarity (0.788→0.862 for Qwen, 0.814→0.874 for Llama) but leaves detectable behavioral differences exploitable by adversaries.

Summary generated by Claude — human-verified