DeepMind's CALLIOPE: LLM-built topic trees beat embeddings under strict privacy
rohanpaul_ai · x · 2026-10-29
A new Google DeepMind paper introduces CALLIOPE: under strict differential privacy, having an LLM build topic trees from keywords yields clearer topic maps than embedding clustering.
Problem: Teams want to know what users ask chatbots without exposing anyone. The usual private approach clusters noisy embeddings, but under strict privacy it returns too few clusters with weak names.
Method: An LLM tags each document with keywords; noise is added so only common keywords survive; the LLM then proposes topics and sorts documents into them.
Results: At the strictest privacy setting tested, human raters preferred CALLIOPE over the embedding-based URANIA on all 3 datasets, by up to 76% vs 18%.
Cost: The method uses 2–5x more tokens — budget accordingly for privacy-safe usage analytics.
Related event: DeepMind's CALLIOPE Enables Privacy-Safe Topic Analysis Without Embeddings(2 posts)→
More from Research
- Prove2Me formalizes 2,000+ operations research papers in Lean 4 with Claude agents, 64% done — ChrSzegedy · 2026-10-30
- 1-16 Coding Agents Tested: Teams Burn 36% More Tokens to Save 5 Hours — kyang8093 · 2026-10-30
- Noninvasive BCI lets unresponsive ICU patients communicate via brainwaves — ZeroStateReflex · 2026-10-30
- Adobe's EditAgent routes image edits across editors with evidence-based listwise judging — adobe · 2026-10-30
- AskStream benchmark shows models struggle to ask questions proactively in streaming video — CASIA · 2026-10-30
- Co-Design Bench: top coding agents fully satisfy only 26.4% of real-user tasks — Qwen · 2026-10-30