Embedding-free framework enables private topic discovery in LLM conversations

CALLIOPE uses LLMs directly to abstract documents, synthesize schemas, and classify text, outperforming embedding-based methods under differential privacy.

Big Tech
Edith Cohen · Vadym Doroshenko · Badih Ghazi · Charlie Harrison · Pritish Kamath · Alexander Knop · +8 more

Google Research · Google DeepMind

Research Digest··3 min read
, LLM conversations) while supporting differential privacy.

The authors designed CALLIOPE, a top-down framework that uses LLMs to recursively partition documents into topic hierarchies.

§