Google Research and Google DeepMind have introduced CALLIOPE, a framework designed to generate hierarchical topic maps from conversation logs while maintaining formal differential privacy guarantees.
The system operates in the discrete text domain, utilizing large language model reasoning to summarize and categorize text instead of relying on traditional vector embeddings.
The process involves three stages: abstraction, where an LLM extracts keywords and summaries; synthesis of a topic schema; and document assignment to specific topics.
According to the researchers, CALLIOPE outperforms its predecessor, URANIA, particularly under strict privacy settings where noise typically degrades cluster quality.
The framework is claimed to perform comparably to or better than CLIO, a non-private baseline system introduced by Anthropic in 2024, while providing mathematical privacy proofs.
