Adaptive Text Anonymization: Learning Privacy-Utility Trade-offs via Prompt Optimization

Damien Riquet; Damien Sileo; Gabriel Loiseau; Marc Tommasi; Maxime Meyer

arxiv: 2602.20743 · v2 · submitted 2026-02-24 · 💻 cs.CL

Adaptive Text Anonymization: Learning Privacy-Utility Trade-offs via Prompt Optimization

Gabriel Loiseau , Damien Sileo , Damien Riquet , Maxime Meyer , Marc Tommasi This is my paper

Pith reviewed 2026-05-15 20:10 UTC · model grok-4.3

classification 💻 cs.CL

keywords adaptive anonymizationprompt optimizationprivacy-utility trade-offlanguage modelstext anonymizationbenchmarkopen-source LLMs

0 comments

The pith

Prompt optimization enables adaptive text anonymization with improved privacy-utility trade-offs

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Text anonymization requires balancing privacy protection with preserving the data's usefulness, but the right balance depends on the specific domain and goals. Static methods cannot adjust to these varying needs and often underperform. This paper formulates adaptive text anonymization as a task solved by optimizing prompts that instruct language models on how to anonymize text for given requirements. A benchmark covering five datasets with different domains and objectives demonstrates that the optimized prompts deliver better trade-offs than prior approaches. The framework operates efficiently on open-source models and can find new anonymization techniques.

Core claim

The paper's central claim is that a framework using task-specific prompt optimization automatically constructs anonymization instructions for language models. This allows the anonymization process to adapt to different privacy goals, domains, and downstream applications, resulting in consistently superior privacy-utility trade-offs compared to static baselines across evaluated settings.

What carries the argument

Task-specific prompt optimization that generates customized anonymization instructions for language models.

Load-bearing premise

That the prompt optimization process discovers anonymization strategies which generalize to new data without overfitting to the benchmark datasets.

What would settle it

Observing that the framework does not achieve better trade-offs than baselines when applied to a new dataset or domain not included in the original benchmark would falsify the general applicability claim.

read the original abstract

Anonymizing textual documents is a highly context-sensitive problem: the appropriate balance between privacy protection and utility preservation varies with the data domain, privacy objectives, and downstream application. However, existing anonymization methods rely on static, manually designed strategies that lack the flexibility to adjust to diverse requirements and often fail to generalize across domains. We introduce adaptive text anonymization, a new task formulation in which anonymization strategies are automatically adapted to specific privacy-utility requirements. We propose a framework for task-specific prompt optimization that automatically constructs anonymization instructions for language models, enabling adaptation to different privacy goals, domains, and downstream usage patterns. To evaluate our approach, we present a benchmark spanning five datasets with diverse domains, privacy constraints, and utility objectives. Across all evaluated settings, our framework consistently achieves a better privacy-utility trade-off than existing baselines, while remaining computationally efficient and effective on open-source language models, with performance comparable to larger closed-source models. Additionally, we show that our method can discover novel anonymization strategies that explore different points along the privacy-utility trade-off frontier.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper frames anonymization as a prompt-optimization task and tests it on five datasets, but the abstract supplies no numbers so the claimed gains are hard to judge.

read the letter

Hi, the core move here is recasting text anonymization as an adaptive problem where prompt optimization finds instructions that balance privacy and utility for a given domain and downstream task. That is a step past the static manual rules cited in prior work, and the five-dataset benchmark is a reasonable way to show the idea in action across domains. They also report that the resulting prompts run efficiently on open-source models and reach performance close to larger closed ones, which is a practical plus if it holds. The claim that the method discovers new points on the trade-off frontier is the part that could matter most for people who need tunable anonymization rather than one-size-fits-all redaction. On the soft side, the abstract gives no quantitative results, no error bars, no dataset sizes, and no description of how the privacy and utility oracles are actually implemented, so it is impossible to tell whether the reported outperformance is large enough to matter or just within noise. The stress-test point about overfitting is worth taking seriously: because the optimization loop scores prompts on the same benchmarks used for final reporting, any advantage could be an artifact of fitting to those particular metrics and tasks rather than a general strategy. Without held-out domains or cross-dataset transfer results, the generalization story stays unproven. This is the sort of paper that would interest NLP people working on privacy-preserving pipelines for clinical or legal text. A reader who already runs prompt-based systems could pick up the benchmark and the optimization recipe and try it quickly. The idea is clear enough and the benchmark adds a concrete testbed, so it deserves a serious referee to check the missing implementation details and run the necessary robustness checks.

Referee Report

2 major / 2 minor

Summary. The paper introduces adaptive text anonymization as a new task formulation and proposes a framework for task-specific prompt optimization that automatically constructs anonymization instructions for language models. It presents a benchmark spanning five datasets with diverse domains, privacy constraints, and utility objectives, claiming that the method consistently achieves a better privacy-utility trade-off than existing baselines, remains computationally efficient, performs effectively on open-source models with results comparable to larger closed-source models, and can discover novel anonymization strategies.

Significance. If the empirical results hold under rigorous validation, the work could meaningfully advance privacy-preserving text processing by replacing static manual anonymization strategies with automated, adaptable prompt-based methods that flexibly balance privacy and utility across domains and downstream tasks.

major comments (2)

[Abstract] Abstract: the claim of consistent outperformance across all evaluated settings supplies no quantitative metrics, error bars, dataset statistics, or ablation results, preventing assessment of the magnitude or reliability of the reported privacy-utility gains.
[Evaluation] Evaluation section: prompt optimization scores candidate prompts directly on the same five benchmark datasets used for final reporting, with no held-out domains, cross-dataset transfer experiments, or explicit regularization against overfitting to the chosen privacy and utility oracles.

minor comments (2)

[Method] The description of the prompt optimization procedure would be clearer with pseudocode or an explicit algorithm box outlining the search loop, scoring functions, and stopping criteria.
[Experiments] Figure captions for any privacy-utility frontier plots should explicitly state the axes, the number of runs, and whether error bands represent standard deviation or confidence intervals.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our work. We address each major comment below with clarifications and indicate planned revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of consistent outperformance across all evaluated settings supplies no quantitative metrics, error bars, dataset statistics, or ablation results, preventing assessment of the magnitude or reliability of the reported privacy-utility gains.

Authors: We agree that the abstract would benefit from quantitative details to better convey the scale and reliability of results. In the revised manuscript, we will update the abstract to include specific metrics such as average privacy-utility trade-off improvements (with standard deviations) across the five datasets, while referencing key dataset statistics from Table 1 and noting that ablation results appear in Section 4.3. revision: yes
Referee: [Evaluation] Evaluation section: prompt optimization scores candidate prompts directly on the same five benchmark datasets used for final reporting, with no held-out domains, cross-dataset transfer experiments, or explicit regularization against overfitting to the chosen privacy and utility oracles.

Authors: This is a fair point on evaluation rigor. Within each dataset, prompt scoring already uses an internal validation split to provide some regularization against overfitting. In the revision, we will add explicit cross-dataset transfer experiments (optimizing on four datasets and evaluating on the held-out fifth) and include a dedicated held-out domain analysis to further demonstrate generalization beyond the optimization sets. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical framework with benchmark comparisons

full rationale

The paper defines a new task of adaptive text anonymization and proposes a prompt-optimization framework evaluated empirically across five datasets. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. Central claims rest on direct performance comparisons to baselines rather than any reduction of outputs to inputs by construction. This constitutes a standard self-contained empirical contribution with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that language models can be guided via optimized prompts to perform effective anonymization; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)

domain assumption Language models can follow optimized prompts to perform context-sensitive anonymization that balances privacy and utility
Core premise enabling the framework; invoked throughout the abstract description of the method.

pith-pipeline@v0.9.0 · 5495 in / 1073 out tokens · 33991 ms · 2026-05-15T20:10:29.987881+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We propose a framework for task-specific prompt optimization that automatically constructs anonymization instructions for language models... evolutionary prompt optimization... GEPA algorithm
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Across all evaluated settings, our framework consistently achieves a better privacy-utility trade-off than existing baselines

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.