acceptodds
Under review as a conference paper at ICLR 2027

SeqBench: Benchmarking Cumlulative Privacy Leakage In Longitudinal Text

Abstract

Language models infer a writer’s attributes from evidence accumulated across posts and linked accounts. We call a target cumulative when the history recovers it and no post alone does. We prove that a per-post anonymiser that detects no more than an evidence-based reader and edits only what it detects leaves every cumulative target recoverable, and that, for a reader of the current archive, only revising released posts repairs a released leak. An anonymiser that edits only what it detects must therefore be sequential: it reads the history and revises what it has released. SeqBench is built to measure this separation. Its eleven synthetic corpora of up to 1,000 personas spread designed targets over several posts, most across platforms, and a blind leave-one-post-out certifier admits a target as cumulative only if a reader recovers it from its support and from no labelled post alone. On six further corpora from the same generator, labelled one post at a time, which mark no conjunctive route, the only kind per-post anonymisation provably cannot see, our sequential anonymiser edits 406 posts that per-post leaves unedited (86 the other way) and reaches back to 282 of its edited posts after release. Under a same-family attacker both cut strict value recall from 0.347 to 0.049 (per-post) and 0.056 (sequential), with no significant difference (p = 0.30); the 90% interval of their type-F1 difference lies within ±0.031. In a four-seat loop run by one minimal-edit script, the sequential scope edits fewer posts and keeps more utility in every seat (+0.058, p = 0.002), while round-4 type recall does not differ significantly.

Primary area:
datasets and benchmarks
Keywords:
text anonymization, privacy, large language models, attribute inference, longitudinal text, cumulative leakage, synthetic benchmark, evaluation methodology

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_n9GTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues that, when a user posts different pieces of text across platforms over time, per post anonymization techniques are insufficient as leakage of certain attributes can happen when evidence accumulates across posts.

Authors provide the following contributions:

  • They define cumulative leakage and show that per post anonymization is insufficient;
  • They propose a sequential anonymizer that rewrites the released posts if the addition of the new post leads to successful inference, following the iterative approach of Staab et al. (2025).
  • They construct a benchmark SeqBench that specifically spreads information about target attributes across posts (1000 generated personas with median of 7 posts), and analyse the performance of sequential anonymization in the utility-leakage trade off.
Critical strengths
  • The paper studies an important problem that, to my knowledge, has not been widely studied. Attribute leakage of text across longitudinal, cross platform posts is highly relevant, and currently not well addressed by text anonymizers or studied in related work.
  • Authors create a synthetic benchmark to measure such leakage, and position their work in related to other benchmarks in the field (Figure 2).
  • They use the state of the art iterative anonymizer as per post anonymizer, so that's a fair comparison.
Critical weaknesses

Overall, the paper does not read well. Sentences often take a long time to understand (very often too short or dense to make things clear), and are not thoughtfully connected. For instance, the paragraph starting online 171: I really struggle to understand what is evidence based or evidence monotone.

It personally really reads as mostly AI generated text (e.g., "a breadcrumb is the atom of the ground truth…" on line 246). While I do not have evidence for that, nor do I believe that's a reason in itself not to accept a paper, the fact that it does not read well remains a major weakness.

  • The results do not compellingly show the benefits of sequential anonymization over per-post anonymization. If I understand it correctly, both the leakage and utility is within the same CI for both in the main results in Table 1. To me, this means that there is no additional benefit from using sequential anonymization in a dataset specifically designed to test leakage that accumulates over different posts, or how should I read this?
  • Fundamentally, the sequential anonymizer assumes that previously released posts can be edited, which is an unfair assumption in practice. An real attacker surely can save original posts before they are edited, and still make attribute inferences based on the evidence accumulated in these.
  • It's tough to validate how realistic the benchmark is, i.e., how it compares to attributes being leaked across posts in reality.

Smaller concerns:

  • Figure 1 can be much better illustrated of accompanied by a better caption, it took me a while to understand it.
  • I do not understand Figure 7.
  • The appendix has a lot of information, but from just reading the main text I had a hard time understanding which specific pieces provide additional details or results for which arguments in the main text.
Questions for authors

Grammar/spelling/structure:

  • Line 56: no single post need give
  • Line 90: you provide a bullet point overview (I guess of the contributions), but no reference of what is listed.

Other questions see weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer n9GT (1/2)

Thank you for the careful reading and for the concrete pointers. Line numbers refer to the submitted PDF.

W1. The paper does not read well (l. 171, l. 246).

We agree. The revision adds a plain-language reading of the reader assumption, defines each term where it is first used, and cuts statements that were made more than once (general response). What the two passages mean:

LineTermPlain meaning
171reader DThe adversary, modelled as something that reads posts and infers values
171evidence-basedD infers a value from phrases one can point to, and still infers it as long as those phrases are in the text
171evidence-monotoneIt follows: giving D more posts never makes it drop an inference. The propositions need this because they argue "the phrases survived the edit, so the inference survived"
246breadcrumbOne planned clue: an indirect fact placed in one post ("trams going past the school gates"). Its record holds the instruction given to the writing model and the attributes the clue helps reveal. It is the unit of the answer key

On why the text reads this way: to fit nine pages, an editing pass put the conclusion first in each paragraph and cut repeated statements, and we attached every qualifier to its claim so that no sentence overclaims. That removed the connecting text a first-time reader needs. The AI-use statement discloses that assistants restructured and edited the text; we reviewed it and take responsibility for it.

W2. Leakage and utility are within the same CI in Table 1. How should I read this?

As you did. On these targets the two arms are equivalent in leakage.

SeqBench-6seedPer-postSequentialReading
Strict value recall (raw 0.347)0.0490.056No significant difference (p = 0.30)
Posts edited67.8%84.2%Sequential edits more
Utility over all posts: surface / semantic / factual0.887 / 0.914 / 0.9560.864 / 0.907 / 0.921Per-post keeps 0.007–0.035 more
Utility in the four-attacker loopSequential edits fewer posts and keeps 0.058 more (p = 0.002)

Table 1 measures cost: whether a defence that reads the history and returns to released posts leaks more or damages more text. It cannot show a benefit, because it runs on six further corpora whose labels were assigned one post at a time (l. 079–081, 397–400), so every labelled clue is visible to both arms. SeqBench's designed multi-post targets are not in it. Section 3 (l. 216–221) says in advance that on such targets Proposition 1 predicts nothing and that "a null ... is uninformative, and so is a win". We wrote the result as a null for that reason.

The benefit claim rests on Proposition 1 and on how widely SeqBench's clue sets are spread: at 1,000 personas every designed set is wider than two posts, 79.3% are wider than four and 35.8% wider than eight (l. 380–384; App. C), so a per-post anonymiser never sees one whole. The comparison on certified cumulative targets is not in the submission, and Section 7 says the reported null cannot test the proposition.

W3. A real attacker can save original posts before they are edited.

Yes, and the paper says so (App. B, "A snapshot adversary"; Section 7, l. 469–471). An attacker who saves everything is the realistic scope of Definition 3, and for it Proposition 2(b) agrees with you: a released leak stays released. The abstract limits the revision claim to "a reader of the current archive".

We kept revision in the paper because sequential anonymisation has two parts, and only one of them changes a published post.

PartChanges a published post?What it doesIn our run (the 406 posts only sequential edits)
Reading the historyNoEdits the newest post before it is released, so the last clue of a set need never go out. This is what Proposition 1 concerns346 edited only at their own step
RevisionYesRepairs a leak the detector finds late, against a reader who arrives after the edit60 reached after release
AuthorsOpenReview ↗

Response to Reviewer n9GT (2/2)

W4. It is hard to validate how realistic the benchmark is.

That is fair, and Section 7 says we ran no human study. SeqBench is a controlled instrument, as SynthPAI is for single-post inference. It tests whether a defence handles cumulative leakage when it is present, and makes no estimate of how common it is in the wild.

QuestionAnswerSource
Why not real data?A cumulative target can only be scored with an answer key that says which posts jointly carry a value. Real corpora have none, and building one would need consented linkage of real people's accounts across platformsl. 226–230
Does the phenomenon exist in real text?Prior work shows it: inference from a handful of ordinary posts; user-level cross-post leakage on social media; identifiers accumulating across one patient's recordsStaab et al., 2024; SopriBench; i2b2/UTHealth
What keeps the corpus coherent?One identity drives style, beliefs, posting rhythm and accounts together, events come from real news headlines, and the population is deliberately confusableApp. E
What stops a clue being too blunt?The certifier does not admit a target as cumulative when one post gives the value awaySec. 4.2, App. F

Figure 1. / Figure 7.

FigurePanelWhat it shows
1LeftOne fabricated person and the channels the generator fixes
1MiddleFour of that person's posts over six simulated days
1RightWhat an attacker concludes. Each post rules out some cities, only all four together leave Manchester, and dropping any one reopens the set
7LeftFor a defence that reads only the last k posts, the share of clue sets too wide to fit in its view (35.8% at k = 8)
7RightThe share of clue sets whose last clue has appeared by the author's D-th post (7.9% by post 20)

Both panels of Figure 7 come from the generator's records, with no attacker involved. The revision says this in the text where Figure 1 is first cited and in the caption of Figure 7.

Which appendix supports which argument?

Main textAppendix
Related resources (Fig. 2); every termA; glossary in Tables 2–3
Edit scopes and the limits of the model (Sec. 3)B
Proofs, the window result, span counts (Sec. 3, 6.1)C
The anonymisation loop and its strength (Sec. 5)D
Personas, corpora, sampling (Sec. 4)E, Table 4
Certifier probes and rules (Sec. 4.2)F
Metrics and utility tiers (Sec. 6)G
Edit sets and editing strength (Sec. 6.2)H
Other attackers and corpora (Sec. 6.2)I
Reproducibility, ethics, release, datasheetJ–M
Model versionsN

The letters above are those of the submitted PDF. The revision folds the fourteen appendix sections into five (A related work and terms, B the formal model, C the benchmark, D the evaluation, E reproducibility and release) and puts this guide, in the new lettering, at the head of the appendix.

Line 56, line 90.

Line 56, "No single post need give an attribute away", means that no single post has to reveal the attribute; the revision uses the plainer wording. The bullets at line 90 are the contributions; the revision adds the lead-in "Our contributions are:".

What we ask you to weigh. You describe the problem as "highly relevant, and currently not well addressed by text anonymizers or studied in related work". The paper's contribution to it is the benchmark: a definition of leakage that no single post carries, a test that admits a target only against a concrete reader, and corpora in which the supporting posts are known. None of the six closest resources labels leakage post by post for one person across platforms (l. 071–073). The anonymiser comparison is a cost measurement, reported as a null. We hope the revised presentation lets the benchmark be judged on those terms.

Reviewer_n9GTOpenReview ↗

Response to rebuttal

Many thanks for the detailed rebuttal. I have read the response and the other reviews/rebuttals, and believe the paper still needs substantial work to be accepted. I keep my score.

Reviewer_SzzpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The work studies cumulative privacy leakage across multiple posts, focusing on information that can be inferred only from a sequence of posts, not from any single post (in practice, many require 8 or more posts). To do this, the authors introduce SeqBench, a new benchmark consisting of eleven synthetic corpora with up to 1,000 people, including cross-platform posts (alongside a few variants). By construction, SeqBench differentiates between the number of posts required for a given disclosure, accepting only synthetic posts that do not reveal an attribute on their own but require other posts to reconstruct it.  The authors then evaluate anonymizing individual posts and sequential anonymization to see whether editing posts independently leaves cumulative leakage unaddressed, which seems to be partly the case (see below).

Critical strengths

Cumulative privacy leakage is a relevant problem with practical implications. Existing work has paid comparatively little attention to larger longitudinal settings, even though sequential release has been studied for earlier privacy notions. A benchmark that allows this problem to be evaluated is therefore a useful direction.

Critical weaknesses
  1. For the proposed dataset (the main focus), the paper provides no information on quality, offers insufficient detail on the generation process (Section 4 is very short and provides little information), and includes no qualitative ablations. Importantly,  after reading the description (and the rest of the paper), I do not feel confident that I could explain either the generation process or the resulting text quality. The idea of splitting attributes into pieces that jointly allow identification is understandable, but the description just names many components without clearly explaining how most are implemented.
  2. Given that, the empirical results do not provide a convincing case for sequential anonymization (in my reading). In Table 1, the difference between sequential anonymization and anonymization of individual posts is only 0.006 F1, while the utility reduction is much more noticeable. The dataset construction also lacks substantial implementation detail.
  3. The assumptions underlying the anonymization setting are also not fully convincing to me. Proposition 1 does not seem to hold for an anonymizer that does more than the minimum needed for each individual post. For example, an anonymizer with access to the user's full attributes could also remove potential hints about these attributes (even in the toy examples given in the paper). Further, the sequential anonymization setup in Figure 6 seems to assume that previously published posts can be edited retrospectively. While the posts themselves could be changed, the usual assumption in privacy work is that information published once is already available to the adversary. This makes the evaluated setup seem unrealistic.
  4. The overall presentation and writing are quite bad. I know the authors acknowledge heavy use of AI, but parts of the work were extremely hard to read and difficult to follow (if at all). This includes (but is not limited to) strange and undefined terminology, unclear figures, and their lack of descriptions (e.g., Figure 2), a barely parseable results section, and generally “atypical” sentences for this type of venue. Somewhat independetly the paper introduces a lot of notation and theory for a rather simple approach (Section 3).
Questions for authors
  1. The paper states that every result about D rests on one assumption (line 171). Which assumption is this?

  2. Does the stated monotonicity of D assume perfect inference? How should this claim be understood for the inference procedure used in the evaluation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Szzp (1/3)

Thank you for the detailed review. You write that a benchmark that "allows this problem to be evaluated is therefore a useful direction". That benchmark is the paper's main contribution, and this first reply says what it contains. Line and section numbers refer to the submitted PDF.

W1. No information on quality, insufficient detail on the generation process.

Section 4 is short because of the page limit, and we agree it names stages without explaining them. The detail is in App. E (personas, corpora, sampling), App. F (certifier) and App. J/M (reproducibility, datasheet), and the generator's code and configs are in the supplementary material. The five stages of Figure 3:

StageWhat happens
1. PersonaOne fabricated identity: demographics drawn from population marginals, a writing-style profile, a belief network, a weekly posting rhythm, accounts on six platforms and a follower graph (Fig. 9)
2. Disclosure planBefore any text is written, a planner splits a chosen attribute into two or more partial clues. Each narrows the value without fixing it (Fig. 4: "sector and headcount" and "biggest employer in the region"). A clue may appear only after 2–10 further posts by the same author
3. SimulationA clock advances, dated news events move beliefs, and the persona posts and replies in threads
4. WritingGemma-4-31B writes each post. A planned clue enters that post's prompt as an instruction ("say X, never name the employer"), so a label names a fact the model was asked to write
5. ExportEach clue's label lists its sibling posts. A clue set that never completes is kept as a decoy

What the submission has on quality:

CheckWhat it testsRef
Corpus statisticsMeasured from the released filesTable 4, App. E
CertifierThe label-quality test: against a blind reader, does a planned clue set recover the value, and does a single post already do so?Sec. 4.2, App. F
Guessing controlsA no-text probe and another persona's clue set, so a guess is not counted as a recoveryApp. F
Constant-guess tableExcludes 12 of 26 attributes on which one fixed answer is right at least half the timeTable 5
Release gatesName collisions and memorisationApp. K
Human studyNone, and Section 7 says soSec. 7

The revision states the stages and these checks in Section 4.

AuthorsOpenReview ↗

Response to Reviewer Szzp (2/3)

W2. The results do not make a convincing case for sequential anonymisation (0.006 F1 in Table 1).

You read Table 1 correctly: the arms differ by 0.006 F1 and we report "no significant difference". Table 1 answers one question and cannot answer the other.

QuestionCan Table 1 answer it?Why
Does reading the history, and returning to released posts, cost leakage or utility?YesBoth arms run on the same targets
Does sequential anonymisation leak less on multi-post targets?NoIts corpora were labelled one post at a time, so both arms see every labelled clue (l. 397–400)

The paper says the second in two places. Section 3 (l. 216–221) says in advance that only targets with a multi-post route separate the two scopes and that elsewhere "a null ... is uninformative, and so is a win". Section 6.2 (l. 397–400) says a difference on these corpora "cannot test it in either direction". We printed that limit beside the result because presenting an equivalence as support for Proposition 1 would overclaim.

The cost question is real: a defence that reads the history and returns to released posts could leak more or damage more text. What Table 1 and its companions report:

SeqBench-6seedPer-postSequentialPer-post minus sequential
Strict value recall, the primary score (raw 0.347)0.0490.056−0.007, p = 0.30
Type F1 (raw 0.625)0.3710.36590% interval within ±0.031
Posts edited67.8%84.2%sequential edits more
Utility over all posts: surface / semantic / factual0.887 / 0.914 / 0.9560.864 / 0.907 / 0.921+0.023 / +0.007 / +0.035
Utility over edited posts only−0.005 (p = 0.30) / −0.016 (p = 0.003) / +0.029 (p = 0.003)

So sequential does not leak significantly more, and its utility loss over all posts comes from editing more of them: over edited posts only, surface fidelity does not differ, semantic fidelity favours sequential and the factual tier favours per-post. In the four-attacker loop, where one minimal-edit script runs both arms, sequential edits fewer posts and keeps more utility (+0.058, p = 0.002; Table 10).

The case for reading the history does not rest on Table 1. It rests on Proposition 1 and on what Section 6.1 measures. At 1,000 personas every one of SeqBench's 4,375 designed clue sets is wider than two posts, 79.3% are wider than four and 35.8% wider than eight, and the median set completes at post 136 of its author's timeline (l. 380–384; App. C). A per-post anonymiser never sees one whole, and a memory of four posts misses 79.3% of them. The revision gives these shares as a table in Section 6. The direct test is the same comparison on certified cumulative targets. The submission does not contain it, and Section 7 (l. 461–462) states that the reported null cannot test the proposition.

AuthorsOpenReview ↗

Response to Reviewer Szzp (3/3)

W3a. Proposition 1 does not hold for an anonymiser that does more than the minimum.

We agree, and the proposition is stated that way on purpose. Hypotheses (i)–(iii) say exactly "detects no more than the reader does in that post, and edits only what it detects".

AnonymiserHypotheses (i)–(iii)Proposition 1
Finds what to hide from the text and edits only that (Staab et al., 2025: edit as little as possible)HoldApplies: every cumulative target stays recoverable
Knows the user's attributes and removes every hint (your example)Breaks (i)Does not cover it
Over-edits, or releases empty textBrokenDoes not cover it; empty text solves the task in every scope (App. B)
Our own loopNot enforced: a whole-post rewrite can remove evidence the detector missed (l. 317–318)Not guaranteed to apply

The introduction says a per-post anonymiser can close a multi-post route without the history only by breaking the hypotheses, "e.g. by over-editing" (l. 066–068). We scoped Proposition 1 to the minimal-edit family because that is where the state of the art sits. The model has no utility constraint, so we report fidelity beside every leakage number (App. B).

W3b. Editing previously published posts is unrealistic; what is published is available to the adversary.

That adversary is in the model. It is the realistic and online scopes of Definition 3, which App. B describes as "the degenerate case in which every state is assumed logged". For it, Proposition 2(b) says what you say: a released leak stays released whatever is published later. We limit the revision claim accordingly in the abstract ("for a reader of the current archive") and in Section 7 (l. 469–471).

We kept the idealised scope because sequential anonymisation has two parts, and only one of them changes a published post.

PartChanges a published post?What it doesIn our run (the 406 posts only sequential edits)
Reading the historyNoEdits the newest post before release, so a multi-post route need not be completed. This is the part Proposition 1 is about346 edited only at their own step
RevisionYesRepairs a leak the detector finds late, for a reader who arrives after the edit60 reached after release

A reader of the current archive is also the adversary of the attribute-inference work we build on: a model shown a user's profile as it stands when the attack runs (Staab et al., 2024). Platforms let users edit and delete, and a later reader sees the edited state.

W4. Presentation; much notation and theory for a simple approach (Section 3).

We accept the criticism of the writing; the general response lists what the revision changes. The anonymiser is simple, and the notation in Section 3 is there to make "cumulative" a property one can test:

NotationWhy the definition needs it
A reader DThe class depends on who reads
Minimal supportsTo say "no single post suffices"
SpanHow much history a defence must see

The certifier of Section 4.2 implements Definition 2 directly, which is why the formal version is in the main text.

Q1. Which assumption does every result about D rest on (line 171)?

Q2. Does the monotonicity of D assume perfect inference?

TermMeaningWhat it does not assume
Evidence-based (l. 171–173)Everything D infers is carried by identifiable phrases, and D recovers a value from a set of texts exactly when the supporting phrases are still in itThat D is correct or complete
Evidence-monotoneIts consequence: showing D more text never makes it retract an inferencePerfect inference. D is any fixed reader; it may miss values or be wrong

Q1: the assumption is that D is evidence-based. Q2: no. The assumption asks only for consistency. Every class is defined relative to D ("the class is a claim about a reader", l. 283), so the certifier tests targets against concrete reader models and certifies only where all of them agree. For the LLM readers of the evaluation, monotonicity is an idealisation and can fail: a model may answer from one post and abstain on the full set. The certifier flags this case and does not count it as a route (App. F, "Routes and classes"), and App. B lists the deterministic reader as a boundary of the model.

What we ask you to weigh. The benchmark is the contribution. None of the six closest resources labels leakage post by post for one person across platforms (l. 071–073), and SeqBench does so at up to 1,000 personas, with a test that admits a target only against a concrete reader. The anonymiser is a reference defence, and we report its result as a null. We hope the revised presentation lets the benchmark be judged on those terms.

I thank the authors for the lengthy clarification bur remain on my initial points (also after reading the other review+rebuttals).

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

The paper is very hard to follow, and its writing and presentation are not at the standard expected for ICLR. Sentences are overloaded with nested clauses, and key terms (“seat”, “agree rule”, “null rule”, “pool attribute”) are used throughout the main text but defined only in an appendix glossary. The abstract alone shows the problem: it relies on undefined terms such as “four-seat loop” and “same-family attacker”, and its central sentence stacks three subordinate clauses before reaching its subject. I suspect this to be a clear case of AI based paper-writing without any human oversight.

Critical strengths

paper is not in a state to be reviewed for this conference.

Critical weaknesses

paper is not in a state to be reviewed for this conference.

Questions for authors

Please detail the AI usage clearly:

  • ideation and domain
  • figures and plots
  • experiments
  • writing style and paper organisation
1 comment
AuthorsOpenReview ↗

Response to Reviewer RveN

Thank you for taking the time to review the paper. We are sorry it was hard to read, and we agree that the presentation fell short. The general response above says what each section is for and where each claim is supported.

Terms. You are right that several terms were defined only in the glossary (App. A, Tables 2–3). We did that to save space, and we see that it made the main text hard to follow. The revision now gives each meaning at first use:

TermMeaning
seata model filling one role (reader, attacker); "four-seat loop" = the same loop run against four attacker models
same-family attackeran attacker model from the same model family as the anonymiser (DeepSeek-V4-Pro against DeepSeek-V4-Flash)
pool attributeone of the 14 attributes whose best fixed guess is right less than half the time (l. 286–287)
null rulea refusal or an uncredited answer counts as a miss (l. 288)
agree rulea target is certified only if every reader model that audited it certifies it (l. 287–288)

AI usage. Thank you for asking; we set it out under your four headings. The paper's AI-use statement (l. 486–501) covers every row but the first.

HeadingWhat AI didWhat we did
Ideation and domainNone. AI assistants did not propose the problem, the definitions or the propositionsProposed the problem of cumulative leakage, defined the cumulative class and stated the propositions
Figures and plotsAI coding assistants produced all figures. The data plots (Figures 7, 8, 10, 11) are drawn by scripts; Figures 1–6 and 9 are diagramsDirected them. The data plots regenerate from saved outputs, and diagrams are marked "illustrative" where they show an example in place of data
Experiments: codeAI coding assistants wrote most of it: the generator, the anonymisation loop, the certifier, the scorers and the analysis scriptsChecked it by script: one script regenerates every evaluation number from saved outputs, and separately written scorers re-derived the corpus statistics. That re-derivation found the detector failure we report in the reproducibility statement
Experiments: models measuredGemma-4-31B wrote every post; DeepSeek-V4-Flash is the anonymiser; DeepSeek-V4-Pro, Claude Sonnet, Qwen3-32B and Gemma-4-31B are attackers (App. N)Report each model and version in App. N
Writing style and organisationAI assistants restructured and edited the text under our direction. An editing pass to fit nine pages put each conclusion first and cut repetition, which produced the compressed style you describeDecided the claims, directed the editing, reviewed the text and take responsibility for it. The revision is written to be easier to read

The revision. We have tried to address your concern in the revised PDF: every change is in red, the terms above are defined where first used, and the contributions are listed with the section that supports each. We hope it now reads clearly enough for the contribution to be seen: a benchmark for leakage that no single post carries. None of the six closest resources labels such leakage post by post for one person across platforms (l. 071–073), and the other two reviews describe it as "a relevant problem with practical implications" and "currently not well addressed by text anonymizers". If you have time to look at the revised Sections 1 and 3, we would be grateful for your view on the content, and for a note of any passage that is still hard to follow.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers (2/2): the evaluation, and what the revision changes

Table G3: Sections 5–6 use the setting.

QuestionFindingTypeRef
What is the sequential anonymiser?The infer–rewrite–verify loop of Staab et al. (2025) with a running profile and the right to rewrite released posts. The per-post arm is the same loop with the profile clearedDSec. 5, Fig. 6
Does it use the history?It edits 406 posts that per-post leaves alone (86 the other way) and returns to 282 released postsESec. 6.2
What does reading the history cost in leakage, where both arms see the evidence?Nothing significant: strict value recall 0.049 per-post, 0.056 sequential (p = 0.30); type-F1 difference within ±0.031ETables 1, 7
What does it cost in utility?0.007–0.035 on SeqBench-6seed, where it edits more posts; it keeps 0.058 more in the four-attacker loop, where it edits fewerETables 1, 10
Which readings go against us?Several favour per-post; all are reportedEl. 442–456, Table 9
What is not shown?The scope comparison on certified cumulative targets; a human studyn/aSec. 7

How the sections support the title.

StepSectionWhat it establishes
13What has to be measured: the class of a target and its span
24Corpora in which both are known, and the test that admits a target
36.1The span is large: 35.8% of clue sets are wider than 8 posts
45, 6.2A defence can be scored in this setting; what reading the history costs, including the readings that do not favour it

What the revised PDF changes. Presentation only. Every change is in red and the main text is now 10 pages. Section, table and line numbers in these replies still refer to the submitted PDF.

ChangeWhere in the revision
Contributions listed under the lead-in "Our contributions are:", each with the section that supports it; plainer wording at l. 056Section 1
Terms defined where first used (evidence-based, shard, breadcrumb, seat, agree rule, same-family attacker)Sections 1, 3, 4 and 6.2
A plain-language reading of the reader assumption; what Proposition 1 covers; the two parts of sequential anonymisationSection 3
The generation stages and the label-quality checksSection 4
The evaluation expanded with results that were in the appendix: how much of the benchmark a per-post or short-memory defence cannot reach (new Table 1), the four-attacker loop figure, fidelity over edited posts, surface-form survivalSection 6
Metric definitions and test details moved to the appendix to make roomSection 6; Appendix D.2
The panels of Figures 1 and 7 explainedSection 1 text; caption of Figure 7
Statements made more than once cut to one, so each result is given once in the bodySections 1 and 3–7; appendix
The appendix folded from fourteen sections into five, with a guide from main-text claims to appendices at its headAppendix

Claims, experiments and numbers stay as they are.

AuthorsOpenReview ↗

General response to all reviewers (1/2): what each section is for

We thank the three reviewers for their time. All three found the paper hard to follow, and we take that as the main problem. Two reviews call cumulative leakage "a relevant problem with practical implications" (Szzp) and "an important problem", "currently not well addressed by text anonymizers" (n9GT); what they question is how we presented and tested it. Here we state what the title claims, what each section is for, and where each claim is supported. Section, figure, table and line numbers are those of the submitted PDF.

What the title claims.

Words in the titleClaimWhere
Cumulative privacy leakageAn attribute that a reader recovers from a person's history but from no single post is a class of its ownSec. 3
in longitudinal textIt lives in one person's posts over time and across platformsSec. 4, 6.1
BenchmarkingSeqBench supplies targets whose supporting posts are known, and a test that admits a target as cumulative against a concrete readerSec. 4

This is a datasets-and-benchmarks submission and the benchmark is the main contribution. The anonymiser of Section 5 is a reference defence that shows the setting can be acted on.

Table G1: Section 3 defines the class and says why it matters. T = theory, D = design, E = empirical.

QuestionFindingTypeRef
What is cumulative leakage?The history recovers the value, no single post does. Sibling classes: single-post, redundant, mixedTDef. 2
Why a separate class?A per-post anonymiser that edits only what it detects leaves every such target recoverableTProp. 1; proof App. C
Is a short memory enough?Only if a whole support fits inside the windowTProp. 3
What does revising released posts add?It is the only way to repair a leak found after release, and only against a reader of the current archiveTProp. 2, App. B
What can an experiment test?Only targets with a multi-post route separate the two scopes; elsewhere a null or a win is uninformativeTl. 216–221

Table G2: Section 4 builds the benchmark.

QuestionFindingTypeRef
Why synthetic?Real corpora have no per-post answer key and would need consented cross-platform linkageDl. 226–231
How is a multi-post target made?A planner splits one attribute into partial clues, 2–10 of the author's posts apartDSec. 4.1, Fig. 4
How is "cumulative" checked?A blind reader sees each post alone and the posts together; the target is admitted only if the set recovers the value and no single post doesDSec. 4.2, Fig. 5, App. F
How big is it?11 corpora, 50–1,000 personas, up to 142,536 posts, 6 platformsETable 4
How far is the evidence spread?35.8% of clue sets are wider than 8 posts; 67.4% cross platformsESec. 6.1, Fig. 7
What do the closest resources lack?None labels leakage post by post for one person across platforms and tests it against a readerEFig. 2, App. A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.