POLAR-Bench introduces a diagnostic evaluation pitting policy-aware trusted LLM agents against adversarial probe models, finding frontier models withhold over 99% of protected attributes while smaller 1-30B models leak substantially more.
provide more details
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents
POLAR-Bench introduces a diagnostic evaluation pitting policy-aware trusted LLM agents against adversarial probe models, finding frontier models withhold over 99% of protected attributes while smaller 1-30B models leak substantially more.