They designed 55 synthetic tasks across 11 categories of personal information, each requiring one missing attribute.
Privacy guidance combining instructions and metadata labels halves agent confidential access
Biéchy et al. show that while system-level instructions alone have limited effect, adding skill-level intrusiveness labels reduces confidential source access by half when combined.
Research Lab
Lucas Biechy · Cédric Eichler · Héber H. Arcolezi · Nicolas Anciaux
Inria · Université Paris-Saclay · INSA CVL · Université d'Orléans · ÉTS Montréal
Research Digest··2 min read
Thread:Agent Rule Compliance
The authors introduce PrivacySkills, a framework with 55 synthetic tasks and 169 skills to evaluate how LLM agents choose among equivalent acquisition pathways (public, confidential, or user interaction).
Why this paper
From Inria and 4 others · Part of Agent Rule Compliance, now 23 papers
In one line
Combining system and skill-level privacy guidance roughly halves LLM agents' use of confidential sources.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§