Paper proposes graduated disclosure for mixed-sensitivity LLM answers
Original title: Contextual Candour: Role-Aware Graduated Disclosure for LLMs over Mixed-Sensitivity Contexts
A reward-based generator cuts unauthorized leakage, including paraphrases, while keeping utility on enterprise and clinical tests.
Researchers argue that binary allow-or-refuse controls over-refuse or leak when large language models reason over mixed public and confidential context. They define graduated disclosure, which scales detail to a user's authorization, and train responses with a reward that favors level-appropriate abstractions while penalizing unauthorized facts, paraphrases, and approximations. On enterprise and clinical benchmarks, the method reduced leakage versus prompt-based, gating, and supervised baselines while preserving correctness and utility.
- Baselines compared include prompting, gating, and supervised methods.
- Penalties cover unauthorized facts, paraphrases, and approximations.
- The authors say disclosure control belongs inside generation, not only at the gate.
From the original article
1 min read at arxiv.orgLLMs increasingly operate over mixed-sensitivity contexts, where they must reason over information ranging from public to confidential while disclosing only what the user is authorised to receive.
By Devansh Srivastav · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.