Reproduce “Privacy Reasoning in Ambiguous Contexts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ren Yi, Octavian Suciu, Adria Gascon, Sarah Meiklejohn, Eugene Bagdasarian, Marco Gruteser
Camber tests whether rationale-guided context can improve privacy judgments for personal assistants. In one example, an alumni coordinator may share an alumnus’s graduation year with a new acquaintance, but sharing authority is unclear. On PrivacyLens+, Gemini 2.5 Pro’s consent condition raised F1 from 76.8% to 95.3%; these gains concern synthetic scenarios, not live clarifications.
Key insights
- LLMs exhibit high prompt sensitivity and a tendency to prioritize privacy protection (avoiding false positives) over utility (permitting appropriate sharing) in ambiguous scenarios.
- Context ambiguity is a significant barrier to accurate privacy reasoning, with models showing less consistency and higher entropy in "appropriate to share" contexts compared to "inappropriate to share" contexts.
- Clarifying context based on LLMs' own identified implicit assumptions (reasoning-guided expansion) is the most effective strategy for improving both the accuracy and robustness of privacy judgments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.