acceptodds
Under review as a conference paper at ICLR 2027

PRIVASIS ENTERPRISE: Learning Need-to-Know Text Sanitization from the Sharing Context

Abstract

To fix an API bug, an external coding assistant needs the stack trace, not the unreleased product described alongside it; an internal release review needs exactly that product. In need-to-know sanitization, a model rewrites records for a recipient and purpose, keeping what the purpose requires and withholding what the recipient should not see. Existing benchmarks spell out what to remove or keep, never testing whether a model can infer this from the request or reverse it for a new recipient. We build , a benchmark and training corpus of synthetic business records whose fact ledger tracks what can be inferred, so indirect leaks count. Paired recipients of the same records need opposite treatment of a fact, and Swap credits a model only when it gets both right. With category-level directives, the strongest general-purpose models look safer when judged by leakage alone, but most of their leak-free outputs also delete the necessary facts from the recipient. Compact sanitizers trained on our corpus meet every requirement more often than these models. Without directives, almost no output from any system meets every requirement, and Swap drops for every general-purpose model. Training on our corpus improves every metric but leaves this gap open, and its per-example objective never compares paired rewrites. Most missed swaps give both recipients the same answer: current sanitizers appear to judge what is sensitive, not who needs to know.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Need-to-know confidentiality, Contextual privacy, Text sanitization, Selective disclosure, Synthetic data, LLM safety, Dataset and benchmark construction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KgJxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies what the authors call need-to-know sanitization. A model receives business records plus a sharing context, meaning a recipient and a purpose. It must rewrite the records so that facts the purpose needs are kept and facts the recipient should not see are removed. The main point is that the same fact can be required for one recipient and forbidden for another.

To study this, the authors build PRIVASIS ENTERPRISE, a synthetic corpus of business records. Each case starts from a fictional event and a fact ledger. The ledger also stores derived facts, each computed from two base facts, so indirect leaks can be detected. Two or three records are written from different employees’ points of view. The same records are then paired with a higher-need and a lower-need context, and an LLM labels each fact as PROTECT, RETAIN or OPTIONAL. The release has 9,000 training contexts with verified reference rewrites and 294 held-out evaluation contexts.

Critical strengths

The problem formulation is the strongest part of the paper. Earlier sanitization benchmarks such as TAB, NAP² and Privasis fix what must be removed in advance. Contextual integrity benchmarks such as ConfAIde, PrivacyLens and CI-Work vary the context, but they do not ask a model to rewrite one fixed set of records in two opposite ways. The strict-swap pairing makes this testable. It separates a model that understands the recipient from one that just removes “sensitive-looking” content. The SWAP metric follows directly from this design and is easy to interpret. The fact ledger with derived facts is also a good idea. It lets the benchmark catch leaks that happen through arithmetic or logical combination, which span-level annotation cannot capture.

Critical weaknesses

The ground truth is an LLM’s normative judgment. Every PROTECT/RETAIN/OPTIONAL label comes from one gpt-oss-120b call per context, followed by deterministic fixes. In the context-only track, the model is asked to recover these labels without seeing them. A low score can therefore mean that the model disagrees with gpt-oss-120b, not that it fails to reason about need-to-know. Many of these decisions are not clear-cut. In Figure 4, it is not obvious to me that approved overtime hours should be hidden from an insurance claims manager. One could argue that it matters for coverage. The human validation is meant to address this, but it is reported in a way that is hard to assess.

Questions for authors

Please report inter-annotator agreement before discussion for the boundary review. Who were the annotators, and did they see the constructed labels during the review? How many labels did they disagree on initially, and how were these resolved?

Reviewer_ZgtdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a benchmark (named PRIVASIS ENTERPRISE), which asks models to rewrite documents based on the recipient and sharing purpose. It checks whether protected information is removed and required information is kept. Paired contexts test whether models handle the same fact differently for different recipients.

Critical strengths

Keeping the documents fixed while changing the sharing context tests whether models adjust disclosure to the recipient.

Critical weaknesses
  • The three annotators agreed with the final labels after discussion, but the paper does not report agreement before discussion, common disagreements, or annotator backgrounds. Clearer policy grounds are also needed for recipient-specific restrictions. Otherwise, it is hard to tell whether the labels follow consistently from the inputs or depend on shared assumptions.

  • The authors' attempts to assess the model's ability to perform this task have limitations. The paper compares instructions with and without category guidance, but does not report prompt comparisons, demonstrations, separate planning and rewriting, or revision after checking. The instruction to keep information when uncertain may also contribute to keeping facts in both contexts.

  • Giving both recipients the same treatment does not distinguish incorrect disclosure judgments from rewriting failures or following the default-preserve rule.

  • Separate tests of fact classification and rewriting with correct labels would clarify the cause. Label distributions and fact counts per context are also needed to interpret the low rate of fully correct outputs.

  • The dataset covers many industries and document types, but all records are synthetic. Its relevance to agents operating under changing conditions therefore remains unclear.

  • Besides, I think the proportions of PROTECT,RETAIN,and OPTIONAL facts, and the distribution of fact counts per context should be reported.

Questions for authors
  • What evidence supports the relevance of this synthetic benchmark to real organizational needs?

  • What policies or guidelines support decisions about what a recipient should not receive?

  • How were the prompts selected? Did you compare alternative prompts, examples, or revision after checking? Does removing the instruction to keep information when uncertain reduce errors where a fact is kept in both contexts?

  • What are the reasons for model failure?

Reviewer_uVovOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies need-to-know sanitization. rewriting the same business records differently depending on who will receive them and why. It introduces synthetic records with fact-level disclosure requirements, including pairs where a fact must be kept for one recipient but hidden from another. The evaluation checks both leakage and loss of needed information. With explicit directives, the trained small models achieve higher full compliance than the general-purpose models highlighted in the paper, despite lower protection-only scores. Without those directives, full compliance is low for every system.

Critical strengths

1.I like the paired-context test. By keeping the record the same and changing only the recipient, the paper tests whether the sanitizer actually considers who needs the information, rather than treating each fact as always safe or always sensitive. 2.The paper also makes a good case for looking beyond leakage. Some models are better at hiding sensitive facts, but they also remove information the recipient needs. The separate human review is more convincing in that comparison.

Critical weaknesses
  1. The benchmark relies on synthetic records and constructed disclosure decisions. The authors validate the evaluation cases, but I’m still unsure how well those decisions reflect real sharing judgments. Since the labels define a correct rewrite, this affects how much I can take from the results.
  2. The context-only setting is the harder and more interesting task, but no model exceeds 3.4% full compliance. That is a useful finding, though it limits the claim that training on this corpus produces a sanitizer that can work out what a recipient needs on its own.
  3. The main idea is clear, but I found the results section a little dense. A short summary of the key takeaway from each setting would make the findings easier to compare.

revised comment (edited on 11/16): on my comment weakness #2, I want to clarify that the paper actually doesn't make the overclaim of context-only competence from training. My concern is more about where the gain happened, because the main positive results are mostly from when directives are given.

Questions for authors
  1. How were the disclosure decisions checked against what people would actually share in these situations? I liked the human validation of the evaluation cases, but I’d like to know where annotators disagreed and whether those disagreements affect the results.
  2. In the context-only setting, what drives the low full-compliance rate? A breakdown would help clarify what the training improves.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.