acceptodds
Under review as a conference paper at ICLR 2027

CINFA: Contextual-Integrity Norm Factorization for LLM Agent Actions via Counterfactual Commit Supervision

Abstract

Contextual privacy in large language model (LLM) agents depends on the recipient, content, and conditions of a proposed disclosure. We present CINFA, an action-conditioned framework for assessing contextual privacy before external actions are executed. CINFA represents each candidate action as a context-action commit that combines available context with its bound recipient, payload, and transmission evidence. Matched counterfactual edits provide global and partially observed factor labels for risk heads trained on frozen LLM representations. The runtime integrates context-guided generation with action checking, factor-guided revision, and re-checking, withholding unresolved actions. Across CI-RL-Synth, PrivacyLens, and ConfAIde, CINFA’s reported average leakage rate is 5.4%, compared with 12.4% for CI-GRPO. Its adjusted leakage rate, measured on helpful outputs, is 7.8% versus 17.3%, while task utility is 2.6 percentage points higher. Component comparisons, controlled factor edits, metadata stress tests, and evaluations across five frozen backbones characterize the evaluated configurations and their runtime inputs. These comparisons assess the integrated system rather than isolate individual mechanisms, positioning action-conditioned checking as a complement to generation-time privacy alignment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Contextual integrity, LLM agents, agent privacy, action-commit learning, latent representations, norm factorization, counterfactual supervision, information-flow privacy

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ocHVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CINFA, a contextual privacy framework that uses counterfactual action examples to train factor-specific linear probes on frozen LLM representations, enabling privacy risk detection and guided revision before agent actions are executed, while reducing privacy leakage and maintaining task utility.

Critical strengths
  1. The use of counterfactual action examples is interesting and well-motivated. Constructing matched examples by modifying different contextual integrity factors (recipient, attribute, and transmission) provides a useful source of fine-grained privacy supervision.
  2. The factor-specific supervision naturally enables factor-guided revision, allowing the system to identify and correct specific privacy violations rather than simply rejecting unsafe actions. This is a practical approach to detect and avoid privacy leakage.
  3. The paper also provides comprehensive evaluations across three benchmarks, including ablation studies, cross-backbone generalization, metadata robustness, and adversarial evaluations.
Critical weaknesses
  1. The individual components, including counterfactual data construction, latent probing, and runtime privacy checking, are closely related to existing approaches. The main contribution appears to be their integration into a unified framework rather than a fundamentally new method.
  2. The method relies on relatively simple linear probes to predict privacy risks from frozen LLM representations. While this simplicity is appealing, it would be interesting to investigate whether richer modeling approaches or more direct uses of factor-specific representations could further improve privacy detection and mitigation.
  3. The counterfactual training data primarily focuses on single-factor violations, while real-world privacy violations may involve interactions among multiple factors. It remains unclear whether the learned factor-specific probes can reliably handle such compositional violations.
  4. The paper would benefit from stronger comparisons with closely related runtime privacy-checking approaches, e.g. Privacy in Action. Moreover, the current ablation does not isolate the benefit of factor-guided revision from the additional factor-specific acceptance checks. A global classifier with generic revision would be an informative baseline.
Questions for authors
  1. Some ablation questions, like how much of CINFA's improvement comes from counterfactual training data versus factor-specific supervision? Would a global classifier trained on the same counterfactual examples achieve comparable performance? And how does factor-guided revision compare with generic revision?
  2. Why are linear probes chosen over more expressive classifiers? Have the authors investigated whether nonlinear probes or other approaches could better capture the interactions here?
  3. How does CINFA compare with Privacy in Action under matched evaluation settings, which I believe might be the more related work here. The authors mentioned this work in the related work part, but unfornaturely didn't compare with it.
Reviewer_77NWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CINFA, a pre-execution privacy checker for LLM agents. The method uses four lightweight classifiers built on a frozen LLM to assess privacy risks associated with each proposed tool call, considering overall risk, recipients, attributes, and information transmission. These classifiers are trained using counterfactual pairs that differ in a single factor. When a potential privacy violation is detected, CINFA attempts to revise the action based on the identified risk and blocks it if the revised action still fails the privacy check.

Critical strengths

Well-motivated choice of inspection point.

Contextual integrity (CI) has become an important research problem in the case of LLM agents. By placing the privacy check after argument binding but before execution, CINFA can in some level evaluate information that earlier-stage approaches may not have access to.

Critical weaknesses

1. The main results cannot be verified as reported, and the effect of withholding on the metrics is unclear.

The headline result (ALR 7.8% vs. 17.3% for CI-GRPO) cannot be verified. ALR is computed only on each method's own helpful outputs, the authors acknowledge that the helpfulness labels and per-run counts are unspecified (A.4), and the paper never reports how often CINFA can withhold actions or how withheld outputs are scored.

2. The checker is not shown to learn contextual norms rather than action priors.

In §3.2 and §4.3, the paper claimed that all matched pairs keep the context fixed and change only the recipient, payload, or transmission condition. So, the heads may never see the same disclosure judged differently under different conditions, which breaks the contextual integrity. They may simply be learning which recipients and payloads look risky. The authors themselves also admit that the comparisons do not isolate the benefit of matched construction.

Questions for authors

Clarification requests (except for critical weaknesses):

  1. For ConfAIde, which tier and subtask are used, and how are instances mapped to candidate actions? Since text-only replies bypass the commit checker, what fraction of ConfAIde outputs actually pass through it?
  2. How does CINFA compare with simply prompting the same Qwen-2.5-7B (or GPT-4o) to judge the serialized commits, or with PrivacyChecker, in leakage, utility, and per-action latency?

Suggestions for future work:

  1. Build a context-flip test set: hold the action fixed and vary only the context, and report head AUROC and flip accuracy on it.
  2. Train the heads on an equal number of unmatched records, and add a lexical baseline (e.g., TF-IDF logistic regression) on the same held-out commits, to isolate what the matched construction and the frozen representations contribute.
  3. Report leakage and utility on a common instance set across all methods, together with CINFA's accept / revise / withhold rates per benchmark, so the ALR comparison has a shared denominator.
Reviewer_Yy71OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CINFA, a privacy checker that inspects LLM agent actions before execution. Each candidate action is processed with its context into an action commit, and matched counterfactual edits of the recipient, payload attribute, or transmission condition provide training labels. Logistic heads on frozen LLM representations predict global and per factor CI risk. At runtime, flagged actions receive one revision guided by the factor heads and a full recheck, or are withheld. Experiments on CI-RL-Synth, PrivacyLens, and ConfAIde show lower leakage and adjusted leakage than CI-GRPO and other defenses, with slightly higher utility.

Critical strengths
  1. The problem studied in this work is critical.

  2. The figures and tables are well-organized.

  3. This work conduct extensive experiments to show the superiority of proposed method.

Critical weaknesses
  1. The motivation for the specific design is unclear. Section 1 motivates checking concrete actions, yet Section 2 mentions that controlled factor variation and latent CI prediction are existing techniques (CI-Steering, PrivGate, NeuroFilter). The paper describes how CINFA differs but does not justify why frozen linear probes with factor heads should be preferred over an existing runtime checker such as PrivacyChecker.

  2. Experiments cannot show the main contribution of the proposed method. Section 4.2 states that the overall comparisons do not isolate commit checking from generation routing or revision. In Section 4.3, every ablation keeps the context probe, and the Full versus Global comparison changes both the acceptance rule and the revision selection, so neither effect is separated.

  3. The paper writing needs substantial improvement. Section 4.3 defines the labels Text and Global twice, for Figures 2c and 2d, without saying whether they denote the same configurations. Section 3.3 uses a “repair trigger” that, unlike the thresholds \taurk\taur_{k}, receives no notation or value in the main text.

Questions for authors

See weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.