acceptodds
Under review as a conference paper at ICLR 2027

FCDC: Federated Contextual Disclosure Control for Heterogeneous LLM Agents

Abstract

Contextual integrity (CI) requires large language model agents to disclose information according to the norms of each interaction. Learning this capability collaboratively is difficult when clients differ both in their disclosure requirements and in the main agents that generate candidate actions. We introduce Federated Contextual Disclosure Control (FCDC), integrating federated CI learning with nested local norm adaptation and shared representation learning. A policy-conditioned contextual encoder provides the common parameter space, while client-specific norm decision modules retain local decision states. A nested objective evaluates shared representations through locally adapted decisions; alternating updates couple local norm adaptation to encoder learning, and only encoder parameters are aggregated. This design supports collaboration across frozen, heterogeneous main agents while keeping interaction records and decision parameters local. Evaluation covers nine clients per fold, two held-out-backbone folds without target adaptation, and action-level and closed-loop privacy–utility outcomes. The tabulated held-out balanced-accuracy gains are approximately 11 percentage points over independent training and 2.6–4.7 points over the strongest federated comparator in each fold. In the MAGPIE-derived comparison with the held-out main agent, FCDC has 5.0 percentage points lower focal private-item leakage and 13.3 points higher constraint satisfaction than an adapted CI supervisor.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Contextual Integrity, Personalized Federated Learning, LLM Agents, Disclosure Control, Nested Learning, Cross-Backbone Transfer

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RqeMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies contextual privacy preservation in federated settings where clients differ in their disclosure norms and the LLM backbones used by their agents. The authors propose Federated Contextual Disclosure Control (FCDC), which combines a shared contextual encoder with client-specific norm decision modules. Through nested alternating optimization, FCDC adapts local decision modules to client-specific norms while collaboratively learning shared representations. Experiments on synthetic contextual integrity data and agent privacy benchmarks demonstrate improved disclosure classification, generalization to unseen agent backbones, and better privacy–utility trade-offs compared to existing baselines.

Critical strengths
  • The consideration of both disclosure norm heterogeneity and main-agent heterogeneity provides an interesting perspective on collaborative privacy learning.

  • The paper compares FCDC against several federated learning baselines and existing privacy protection methods. It also evaluates generalization to unseen LLM backbones and considers both privacy leakage and task utility in agent interactions, providing evidence of the proposed method's effectiveness.

Critical weaknesses
  1. Definition of norm heterogeneity: relation to Contextual integrity? The relationship between norm heterogeneity and contextual integrity could be clarified. The first paragraph of the Introduction motivates contextual integrity through context-dependent disclosure norms, while the second paragraph introduces norm heterogeneity as differences in disclosure conditions, such as recipient verification and consent requirements. A clearer transition between them could clarify how norm heterogeneity arises from contextual integrity.

  2. Definition norm heterogeneity: explicit or implicit norm? Appendix B.1 constructs norm heterogeneity using three predefined policy variants (contextual minimum, verified recipient, and explicit consent), which appear to represent different explicit disclosure requirements rather than implicit variations in contextual norms. Is norm heterogeneity intended to capture differences in contextual privacy expectations across clients, or differences in explicitly specified disclosure policies?

  3. Practical relevance of norm heterogeneity. It is unclear how the norm heterogeneity maps onto realistic deployment scenarios. Could the author provide real-world examples?

  4. The proposed nested learning formulation is inspired by FedRep, which optimizes shared representations through alternating local predictor adaptation and representation updates. A comparison or clarification of technical modifications would further strengthen the work.

Questions for authors
  1. Could the authors clarify how norm heterogeneity relates to contextual integrity? In particular, does the proposed formulation aim to capture differences in contextual privacy expectations across clients, or differences in explicitly specified disclosure policies? How do the three predefined policy variants in Appendix B.1 reflect the broader notion of contextual norms?
  2. Could the authors provide a concrete real-world deployment scenario illustrating how client-specific norm heterogeneity arises in practice? In such a scenario, why would clients need to learn separate norm decision modules rather than relying on a shared policy-conditioned model?
  3. What technical modifications are required to adapt FedRep's alternating local predictor and shared representation learning to the proposed contextual integrity setting?
Reviewer_wBSTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper explores disclosure control in a federated, collaborative setting based on the norms of each interaction. The authors propose a framework that encodes task-relevant information locally for each agent, which is then aggregated together and learned in a federated and collaborative manner. The experimental results demonstrate that the proposed framework improves balanced accuracy over other federated learning algorithms.

Critical strengths
  • The work falls under an important research direction of ensuring LLM agents make contextually appropriate data-sharing decisions.
  • Combining contextual integrity, federated learning, and collaborative learning as the problem setting is unique.
Critical weaknesses

My biggest concern about this paper is the clarity and presentation. After going through the abstract and introduction, I still have no idea what problem the paper is tackling. The paper begins by motivating contextual integrity through a brief example, but then suddenly introduces collaborative and federated learning without motivating why they are needed or relevant to the setup. Even after reading the problem formulation section, the setup is still not clear.

Moreover, the paper introduces a lot of non-standard terms without defining them. For example, what is norm heterogeneity? What is a main-agent? How do you define a norm in this setup?

The presentation of the proposed framework is even more difficult to understand. There is no intuition behind the design choices, and it just feels like arbitrary components compiled together.

Lastly, I have some concerns about the experimental evaluation choices. In particular, the datasets used and the baselines compared against, which are explained in more details in the Questions section.

Questions for authors
  • Why is federated learning needed? I’m going to assume that the norms are what are being collected and aggregated. Generally, federated learning is needed when the data is sensitive, so it can be stored locally rather than being directly accessed by a central server. However, do norms contain sensitive information?
  • What is a CI input, and why did you define it as that five-tuple in Eq. 1? In CI works, people define an information flow using the standard five CI parameters (data type, data subject, data sender, data recipient, and transmission principle). However, this CI input is non-standard and bears no resemblance to the standard CI information flow. What is the connection between these two?
  • Why were the synthetic scenarios from CI-RL, which were originally used for RL training LLMs to improve contextual reasoning, the dataset used for evaluation? Since this work is targeting contextual integrity, I would expect more standard datasets, such as ConfAIde and PrivacyLens, to be used here.
  • Why were only federated learning algorithms compared against? Shouldn’t prior work specifically targeting contextual integrity also be compared?
Reviewer_mHbQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper trains a small classifier (ModernBERT-base plus a linear head) and applies it between an agent and the outside world. The classifier either allows or denies disclosed information. The paper trains this classifier with federated learning.

Critical strengths

None

Critical weaknesses

The paper is not easy to understand and is full of heavy AI jargon. For example, figure 1 does not explain what CFE or NDMs mean. The abstract mentions "MAGPIE-derived comparison" but this is not known at this point. Section 5 mentions "Following Section 3’s transfer protocol, backbone slots A–D form two folds: FD trains on A/B/C and transfers to D; FC trains on A/B/D and transfers to C", section 3 does not explain this. Unclear writing is the main critical reason for my review. The paper's technical details cannot be assessed or followed.

The paper lacks clear motivation. What is the task/policy of each client? Why do we need federated learning? What is the practical usecase? When do clients use one policy at a time? Why don't we train a model that takes the policy and adapts based on the policy dynamically?

Details of policies per client (which is the central contribution of the paper), are unclear. Contextual Integrity Synthetic Dataset released with CI-RL does not support policies by default. The description in Appendix B (Each fold crosses three source backbone slots with three fixed policy texts: contextual minimum permits only task-appropriate information flows; verified recipient additionally requires confirmation of the intended recipient; and explicit consent additionally requires a current receipt scoped to the task and recipient) is unclear. What is an example of the policy text/natural language? The dataset does not also contain fields for trusted vs untrusted context.

Questions for authors

I have no questions that will change my assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.