acceptodds
Under review as a conference paper at ICLR 2027

BRIEF: Bounded-Context Regeneration of Long Documents with Verifiable Structural Topology

Abstract

Long legal and financial documents combine extreme length with checkable dependencies, including numbered clauses, section containment, defined-term reuse, and internal cross-references. We study whether such documents can be regenerated without relying on a single generation prompt whose context grows with document length. We present BRIEF (Bounded Regeneration In Exact Form), a graph-conditioned schedule for bounded-context regeneration. BRIEF extracts a reference graph, builds a target graph, and writes each clause from a short structural brief containing its number path, required terms, and cross-references. The generator never receives source prose or sibling clauses. Across the controlled length sweep, BRIEF maintains cross-reference fidelity near 0.98 while keeping maximum input-prompt length approximately constant at 86 tokens. On CUAD contracts and EDGAR EX-10 exhibits, BRIEF improves extractor-verifiable section and cross-reference topology over one-pass generation. A retrieval-based linkage audit shows lower source linkability after regeneration, although substantial residual leakage remains. A separate coherence analysis shows that preserving this topology does not by itself preserve reference semantics. A bounded, dependency-ordered plug-in, BRIEF-L, restores most of this within-release coherence, from near zero to between 0.44 and 0.70 of cited clauses, while keeping cross-reference fidelity within 0.03 of BRIEF and adding about 0.08 to EDGAR linkage. A cross-generator replication on Llama-3.1-8B preserves the bounded-context advantage and the low-coherence finding, while showing that BRIEF's relative advantage over the naive bounded baseline is generator-dependent. BRIEF does not establish formal anonymity, general semantic equivalence, or preservation of legal and financial meaning.

Primary area:
generative models
Keywords:
Long-context generation, bounded-context generation, structured document regeneration, graph-conditioned generation, long-document generation, reference structure, cross-reference fidelity, privacy-preserving text generation, synthetic text, document de-identification, source linkage, privacy auditing, structure-preserving generation, legal documents, financial documents, LLMs

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_7hwYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The motivation is data scarcity in law and finance: real contracts and filings are useful for training and testing models but contain private information, so the authors want to release synthetic substitutes. Such a substitute should keep the reference structure that downstream tools rely on (clause numbering, section containment, defined terms, internal cross-references) while regenerating the source-specific wording. The research question, as the abstract states it, is whether long documents of this kind can be regenerated without a single generation prompt whose context grows with document length. Privacy is treated as a secondary, empirical audit rather than a guarantee.

BRIEF parses the source into a content graph, sanitizes and canonicalizes it into a target graph, and writes every clause in its own LLM call from a short "brief" (number path, terms to define or use, the exact cross-reference string, plus a topic, an opening steer and a shared register instruction). No call sees source prose or any other clause, so the prompt stays at about 86 tokens at every length of the synthetic sweep. Fidelity is an exact, extractor-based F1 over containment and cross-reference edges keyed by number path, scored against the target graph. The paper also proposes a per-document capacity diagnostic for structural perturbation, and BRIEF-L, which orders clauses so that cited clauses are written first and adds up to k = 4 clipped summaries of the cited clauses to each prompt.

Experiments use a seeded synthetic corpus (70 to 1200 clauses), 70 CUAD contracts and 47 EDGAR EX-10 exhibits, Qwen3-4B and Llama-3.1-8B as generators, Qwen3-14B and Phi-4 as judges, and a BGE retrieval attacker that holds the source corpus. BRIEF keeps cross-reference F1 near 0.98 on the synthetic sweep, where one-pass generation scores 0.0 on the section tree and erratic cross-reference F1, and it lowers BGE top-1 linkage relative to one-pass on CUAD (0.37 vs 0.70) and EDGAR (0.23 vs 0.49). The authors also report that a per-clause baseline with a generic prompt matches or beats BRIEF on topology, that within-release cross-reference coherence under BRIEF is near zero (0.02 CUAD, 0.06 EDGAR), and that BRIEF-L raises it to between 0.44 and 0.70.

Critical strengths
  1. The paper measures a clear dissociation between reference topology and reference semantics. BRIEF's internal cross-references mostly resolve (0.83 on CUAD and 0.88 on EDGAR, Table 13), yet almost none of them point to a clause that makes sense as a target (coherence 0.02 and 0.06, Table 22), and clause roles fall below the majority-class rate (Table 21). Given that the generator never sees the cited clause, the direction is expected, but the size of the gap is a useful warning for anyone who scores generated documents with structural checkers alone. BRIEF-L is a simple, bounded fix (write cited clauses first, pass their summaries forward), and the ablation in Table 23 shows that the summaries, not the edge-consistent topics, carry the gain. This is the part of the paper I learned the most from.

  2. The reporting is careful and candid. The authors keep the naive per-clause arm even though it matches or beats BRIEF on topology (Table 3), score against the raw source graph as well as the sanitized target (Table 15), report paired tests with document-level bootstrap intervals, replicate with a second generator family where the replication only partly holds, and list what is not preserved (Table 9). Most of my concerns below are built from numbers the paper itself provides.

Critical weaknesses
  1. The structural scores do not yet establish useful long-document regeneration. BRIEF prepends number paths in code, so perfect section-tree F1 partly measures construction rather than model capability; a generic per-clause prompt also reaches cross-reference F1 of 0.98-1.00 (Table 6). One-pass section-tree F1 is already zero with a 1,215-token prompt at 70 clauses (Table 7), which does not establish degradation with length and leaves a rendering/extractor mismatch plausible. Cross-reference performance has no monotonic Qwen length trend. A section-level or plan-then-write baseline is missing between one-pass and clause-wise generation. More importantly, clause-role agreement falls below the majority baseline (Table 21), and no downstream or human utility test shows that preserved numbering and reference strings with unrelated clause content serve the motivating applications. The topology/semantics dissociation is informative, but it does not itself validate the release's usefulness.

  2. The topology, coherence and privacy claims refer to different operating points. The headline release keeps one canonical cross-reference per clause; BRIEF-L is evaluated with up to four (Appendix S). In that latter setup even summary-free BRIEF has CUAD linkage of about 0.89, versus the headline 0.37, and BRIEF-L remains at 0.88. The low-linkage headline release has raw-source cross-reference F1 of only 0.67 on CUAD and 0.49 on EDGAR (Table 15). Thus the abstract's small BRIEF-L fidelity/linkage changes are relative to the more-linkable k-citation baseline, not the headline arm. Section 6.7 acknowledges the shared k-citation comparison, but does not make the difference from the headline operating point sufficiently clear. No reported configuration jointly demonstrates preserved source topology, coherent references and low linkage.

Questions for authors
  1. How many internal cross-reference edges does the headline target graph keep, and how many does the k = 4 target keep, relative to the raw source graph? Is the number of emitted cross-reference strings the only difference between the headline BRIEF arm and the summary-free BRIEF arm of Appendix S (same prompts, topics, defined terms and document window)? A rise from 0.37 to 0.89 under a mean-pooled embedding attacker from a few extra "Clause x.y" strings per clause is larger than I would expect. If the cross-reference count is the only difference, I would read the privacy claim as applying to a heavily reduced structure, which changes my view of the central trade-off. If you already have the k = 4 releases, their raw-source cross-reference F1 next to Figure 4 would help.

  2. What exactly does the one-pass arm receive? The Introduction (p. 1) says it "places the whole source in context and rewrites it," while Appendix H and INV-2 (Sec. 3) indicate that its prompt lists the numbered clauses with the per-clause instructions in brackets and contains no source prose. If it does see source prose on CUAD and EDGAR, its higher linkage is partly explained by that input. What was the maximum output length for one-pass, what fraction of one-pass outputs were truncated at each length, and was Llama's 32K budget (Appendix O) binding at 1,200 clauses? The 0.19 in Table 18 could reflect truncation.

  3. For Table 7, what does the extractor see in a one-pass output at 70 clauses that gives section-tree F1 of exactly 0.0 even when the two-level format is stated in the prompt? An example of the raw output would show whether this is a generation failure or a parsing convention.

  4. Appendix I (p. 19) says candidate-only edges "are not folded into the F1," but Eq. (1) has |C| in the denominator, and Appendix L (p. 21) attributes one-pass's low fidelity to over-generation of references "rather than omission." Which definition produced the reported numbers? If invented edges are excluded from C, the F1 behaves like a recall, and I do not see how over-generation would lower it.

  5. Appendix R explains that the copy arm's coherence differs across tables because each table judges a different subsample or citation setup (0.85 on CUAD in Table 22, 0.57 in Table 23 under the same judge, 0.40 under Phi-4). That spread for the same source documents still suggests a noisy probe, and BRIEF-L scores above the source on EDGAR. Was the probe checked against human judgments on any sample? BRIEF-L writes the citing clause from a summary of the cited one, so a judge that rewards topical overlap would score it highly by construction, while real cross-references often connect clauses on different topics.

  6. What does "sanitize" do to defined terms and other strings carried by the target graph? Table 26 shows terms such as "NICE" and "UNDER" that look like source strings or extraction errors. Which source strings are emitted verbatim in a release, and how much of the residual linkage do they account for?

Suggestions for a future version, not needed for my assessment: a deterministic template rendering of the target graph as a baseline; a section-level or plan-then-write baseline in the style of AgentWrite; a larger long-context model for the one-pass arm; and an LLM-based rewriting anonymizer, for example adversarial anonymization (Staab et al., arXiv 2402.13846) or DP-Prompt paraphrasing (Utpala et al., arXiv 2310.16111), as a privacy baseline next to NER masking, which masks only about seven percent of the text (Sec. 6.4).

3 comments
AuthorsOpenReview ↗

Response to Reviewer

Thank you for a careful review; several of your suspicions were right. Setup and shared points are in the general response.

Q1: edges kept, and the 0.37→0.89 linkage gap.

  • Edges. Raw source graphs have 98 (CUAD) and 151 (EDGAR) internal cross-reference edges. The headline k=1 targets keep 81 and 140; the k=4 targets keep all retained edges.
  • Term-uses. The headline and Appendix S arms also differ in the number of term-uses carried per clause (1 vs 4), not only in the cross-reference count. We isolated each factor:
brief carriesCUAD BGEEDGAR BGE
1 xref, 1 term-use0.420.21
4 xrefs, 1 term-use0.430.20
1 xref, 4 term-uses0.890.50
  • Extra "Clause x.y" strings do almost nothing; the jump comes from defined-term names. 47–73% of source defined terms appear verbatim in releases, while fewer than 2.5% of 4-grams do. So the headline privacy result applied to a reduced brief, as you suspected; Sec. 6.4 and the abstract now say so.
  • Pseudonymization. We map each defined term to a generic name from a per-document pool, with positions unchanged. BRIEF-L at k=4 with the full graph then reaches:
    • CUAD: linkage 0.029 (chance 0.014), coherence 0.63, target xref F1 0.85.
    • EDGAR: linkage 0.007 (chance 0.021), coherence 0.63, target xref F1 0.86.
    • On Llama-3.1-8B: linkage 0.024 / 0.014.
  • Raw-source F1 at k=4 (new Tables 29–30, Appendix U), BRIEF / BRIEF-L / naive-L: CUAD 0.75 / 0.68 / 0.68; EDGAR 0.58 / 0.50 / 0.48.

Q2: one-pass input and truncation.

  • One-pass receives the numbered clause list with bracketed per-clause instructions and no source prose. The Introduction's "places the whole source in context" was wrong and is corrected.
  • The output cap scales with length up to 40k tokens (4.9k at 70 clauses, 40k from 600 clauses on), as in the submission. On Qwen, every one-pass output from 70 to 1,200 clauses (5 documents per length) stopped on its own, using a third of the cap up to 600 clauses and at most three quarters at 1,200.
  • On Llama-3.1-8B we reran 600 and 1,200 clauses (5 documents each). At 1,200 all five outputs reach the 40k cap, and three of five do at 600, but only after the last requested section: the model keeps writing invented sections or repeats instead of stopping. The cap removes run-on text, not requested clauses, so a cut-off document does not explain Table 18's 0.19: 952 to 960 of the 960 requested sub-clauses are present in every output. Our rerun gives cross-reference F1 of 0.00 to 0.67 (mean 0.40, five documents), so it does not reproduce 0.19 exactly.

Q3: Table 7 zero.

  • It is a parsing convention, not a generation failure. Up to 900 clauses the model writes each top-level section as an empty heading ("1."), and the extractor creates no node for an empty heading.
  • At 1,200 clauses it omits the 240 headings and writes exactly the 960 sub-clauses, correctly numbered. Either way no containment edge can match.
  • Re-scored with those headings counted as nodes (inferred from sub-clause numbers at 1,200), one-pass on Qwen3-4B reaches section-tree F1 0.93–1.0 at every length (Appendix U, Table 34). So one-pass keeps the hierarchy, and Table 7 now says so.
  • One-pass cross-reference F1 is lower and less stable beyond 70 clauses: 0.64–0.80 at 70 clauses vs 0.02–0.64 at 300–600.

Q4: Eq. (1) vs Appendix I.

  • Eq. (1) produced every reported number, and invented edges are in C, so over-generation lowers F1 through precision.
  • Appendix I wrongly described the empty-target skip as excluding candidate-only edges; it is fixed, with pseudocode and an example (T = {1→3}, C = {1→3, 2→4} gives F1 = 0.67).
  • Precision vs recall for one-pass and BRIEF:
    • CUAD: one-pass 0.42 / 0.86, BRIEF 0.66 / 0.81.
    • EDGAR: one-pass 0.24 / 0.80, BRIEF 0.42 / 0.82.

Q5: coherence probe.

  • Random-target control. The real cited clause scores 0.59–0.71 for BRIEF-L. Replacing it with a random clause from the same release scores 0.08–0.17. So the probe separates real from random targets, and BRIEF-L's gain is more than topical overlap with any clause.
  • Template and source. The deterministic template ("This is subject to Clause x") scores 0.55–0.57. The probe rewards non-contradiction more than specific fit, and the revision states this. The source itself scores 0.56 (CUAD) and 0.38 (EDGAR), below BRIEF-L's 0.63: BRIEF-L writes the citing clause from a summary of the cited one, so the probe favours it by construction. We read it as agreement between citing and cited clause, not realism. The copy arm's spread across tables comes from different subsamples and citation setups, and we do not claim the probe is low-noise.
  • Judges. Qwen3-14B and Phi-4 agree on arm rankings (Spearman 1.0, 1.0, 0.94), although Phi-4's absolute scores are lower.
  • Human labels. We did not run a human study in the rebuttal window. The camera-ready will report agreement between human labels and the judge on blind pairs.

Q6: sanitization and verbatim strings.

  • In prose, 1.2–2.4% of 4-grams and at most 0.1% of 8-grams appear in the source.
  • Defined terms are the leak: 47–73% are emitted verbatim, and they drive linkage (Q1). Pseudonymization removes them.
  • Entity leakage under pseudonymization:
    • ALL-CAPS and organisation names fall to 0–0.4% (from up to 28%).
    • Dates and amounts are about 0%.
  • "NICE" and "UNDER" are term-extraction artefacts. Table 9 now lists defined terms as preserved unless pseudonymized.

W1: usefulness, and the plan-then-write baseline.

  • Assembly. We agree that section-tree F1 = 1.0 comes from code-side assembly and now present it as an invariant. The template renderer confirms it: F1 0.97–0.99, same linkage as BRIEF.
  • One-pass, numbering equalized. BRIEF still has higher CUAD fidelity (0.91 vs 0.82) and, without pseudonyms, lower linkage (0.89 vs 0.96); EDGAR ties (0.92).
  • Rolling window (brief + previous 4 clauses). Target F1 is 0.88 / 0.87, but coherence is 0.21 / 0.22 (near its random control) at about 2× tokens. Graph-conditioned summaries beat an unconditioned window.
  • Downstream. Three LLMs answered "which clause does X cite" on sources and on releases. Their ranking is identical in every condition, and accuracy is higher on releases because references are explicit.
  • We claim only that releases preserve the ranking of systems on this structure-dependent task, not absolute accuracy or legal meaning. A sub-clause-count probe was not informative.

W2: operating points. With pseudonymization, one configuration (BRIEF-L, k=4, full graph) shows topology, coherent references and chance-level text linkage together; the numbers are under Q1.

Structural linkability.

  • A graph-matching attacker on canonicalized sources links about 0.72 of releases for every arm, including the template.
  • A numbering-invariant attacker still links about 0.6 after all sections are merged.
  • Our privacy claim now covers only the release text, which contains no party names or confidential details; the section tree is a quasi-identifier of any release that keeps it.

Suggestions. We added the template baseline and a rolling-window baseline, the closest we ran to plan-then-write (it is not AgentWrite-style planning). We did not run a larger long-context model for one-pass. An LLM anonymizer baseline (DP-Prompt or adversarial anonymization) is discussed in Related Work. We did not run it, since pseudonymization already reaches chance text linkage and structure is the remaining channel.

Dear Reviewer 7hwY ,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

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

The paper studies regeneration of long structured documents using bounded-context language-model calls. BRIEF extracts and canonicalizes a document’s reference graph, then generates each clause from a local structural brief containing numbering, term, and cross-reference requirements. BRIEF-L adds dependency-ordered generation and bounded summaries of cited clauses. The paper also introduces a structural-fidelity metric, an operator-specific perturbation-capacity diagnostic, and a source-linkage audit. Experiments use a synthetic length sweep, CUAD contracts, and EDGAR exhibits with two generator families. The reported results show improved reference-topology fidelity over one-pass generation, reduced linkage under the tested retrieval attacker, and improved within-release reference coherence when cited-clause summaries are provided.

Critical strengths

Separating structural fidelity from reference coherence, together with the summary-conditioning ablations, provides useful evidence that information about cited clauses improves within-release coherence beyond topic alignment alone.

Critical weaknesses
  1. Required cross-reference strings are supplied explicitly, and canonical clause numbers are prepended programmatically rather than generated by the model. Section-hierarchy preservation is therefore largely built into the assembly procedure, rather than evidence that the model maintains long-document structure. Table 3 shows no statistically significant cross-reference-fidelity advantage over the generic bounded baseline in any of the four corpus–generator comparisons, with BRIEF significantly worse in three. Likewise, naive-L has numerically higher coherence than BRIEF-L in every reported row of Table 23, although BRIEF-L achieves lower BGE linkage. These results suggest a trade-off between prompting choices, rather than an unambiguous advantage of the enriched method. A template-based reference renderer and a concrete downstream structural task would help establish what the generated prose contributes beyond reproducing the supplied graph. Without such evidence, the advance beyond scaffolded local prompting and bounded dependency summaries appears limited.

  2. The headline BRIEF experiment retains a single primary cross-reference per clause, whereas the BRIEF-L experiments retain up to kk references. Appendix S identifies this change and reports that the summary-free CUAD baseline’s linkage increases from approximately 0.37 to 0.89 in the latter setting. BRIEF-L itself has CUAD linkage of 0.88 and 0.86 for the two generators. The authors appropriately caution against comparing these settings directly, so the increase cannot be attributed to summaries alone. However, the low-linkage headline result and the improved-coherence result do not jointly demonstrate an operating point with substantial structural coverage, useful coherence, and strong linkage reduction. These quantities need to be evaluated together with the retained target graph and generation budget held fixed.

  3. The privacy audit evaluates text representations and coarse layout features, but not an attacker that matches canonicalized reference graphs. This omission matters under the source-available threat model because the method deliberately preserves reference topology. Appendix J reports unique structural signatures for all 69 CUAD documents and 36 of 47 EDGAR documents. These counts do not establish any particular attack success rate after regeneration, but they make a structure-aware linkage evaluation especially important. The concern is not the absence of formal anonymity or source-semantic equivalence, which the paper explicitly disclaims. It is that reduced retrieval similarity and high fidelity to a restricted target graph do not yet establish useful privacy-oriented document regeneration.

Questions for authors
  1. How many original internal cross-reference edges survive canonicalization and the single-reference or kk-reference restrictions, before generation? Please distinguish preprocessing losses from generation errors and clarify whether the reference cap changes the scoring target or only the information supplied to the generator.

  2. Does CC in Equation (1) include invented edges? Appendix I says candidate-only edges are not folded into F1, whereas the stated equation would penalize them through ∣C∣|C|. Does that exclusion apply only to empty target families, and which convention produced the reported tables?

  3. With the retained target graph and generation budget held fixed, what evidence demonstrates an advantage of BRIEF/BRIEF-L over naive/naive-L or a template-based reference renderer on a concrete downstream task? Reporting topology, coherence, and linkage together in that matched setting would materially affect my assessment.

  4. What source-linkage accuracy is obtained by an attacker that applies the published canonicalization procedure to the source corpus and matches its reference graphs against the released document’s graph?

4 comments
AuthorsOpenReview ↗

Response to Reviewer

Thank you for the careful reading. Q3 asked for a matched setting with topology, coherence and linkage reported together, so we ran it. Setup and shared points are in the general response.

W1: structure built into assembly; BRIEF vs naive. We agree on both points.

  • Assembly. Section-tree F1 = 1.0 is an invariant of code-side assembly, and the revision presents it that way. A deterministic template renderer (no LLM, "This is subject to Clause x") confirms this: target xref F1 is 0.97 (CUAD) and 0.99 (EDGAR), with the same BGE linkage as BRIEF.
  • Prose. Generated prose adds lexical diversity; cited-clause summaries (BRIEF-L) make references coherent.
  • BRIEF vs naive. BRIEF's prompt does not beat naive on topology, and with term pseudonymization (below) both reach chance linkage, so we drop linkage as a BRIEF-vs-naive claim. What remains, at k=4 with pseudonyms:
+ pseudonymizationlinkage CUAD / EDGARcoherence (random-target) CUADcoherence (random-target) EDGARclause bodies repeated verbatim
BRIEF-L0.029 / 0.0070.63 (0.17)0.63 (0.14)0.1% / 0.3%
naive-L0.029 / 0.0000.64 (0.30)0.71 (0.28)12.5% / 14.4%

naive-L repeats one clause body in eight. Its coherence is equal or higher, but so is its random-clause score, so its real-vs-random margin is narrower (0.34 vs 0.46 CUAD; 0.43 vs 0.49 EDGAR). (Without pseudonyms, BRIEF-L also links less at matched k=1: CUAD 0.52 vs 0.65, EDGAR 0.28 vs 0.36.)

W2 / Q3: one operating point with structure, coherence and low linkage, with budget and target held fixed. The headline's low linkage came from carrying one defined-term use per clause, not from the single cross-reference:

brief carriesCUAD BGEEDGAR BGE
1 xref, 1 term-use0.420.21
4 xrefs, 1 term-use0.430.20
1 xref, 4 term-uses0.890.50

47–73% of source defined terms appear verbatim in releases. We add term pseudonymization to the method: before generation, each defined term is mapped to a generic name from a per-document pool, while definition and use positions stay the same. At k=4 with the full retained graph:

BRIEF-L + pseudonymizationBGE (chance)coherence (random-target)target xref F1tokens
CUAD0.029 (0.014)0.63 (0.17)0.85 (was 0.88)8.9k
EDGAR0.007 (0.021)0.63 (0.14)0.86 (was 0.89)13.3k

On Llama-3.1-8B: 0.024 / 0.014 linkage, coherence 0.50 / 0.58.

Q3: downstream task.

  • Qwen3-4B, Llama-3.1-8B and Qwen3-8B answered 125 "which other clause does clause X cite?" questions (gold from the canonical source graph) on sources, BRIEF, BRIEF-L, BRIEF-L + pseudonymization and the template.
  • The model ranking is identical in every condition (Llama > Qwen3-4B > Qwen3-8B); accuracy is higher on releases (0.45–0.77 vs 0.29–0.44), where references are explicit.
  • We claim only that releases preserve system ranking on this task. A sub-clause-count probe was not informative.

W3 / Q4: graph-matching attacker.

  • Attacker. We applied the published canonicalization to every source and matched reference graphs to each release (edge Jaccard plus a numbering tie-break). It links 0.48 (CUAD) and 0.74 (EDGAR) at the headline setting, and about 0.71–0.76 for every arm at k=4, including the no-LLM template. Pseudonymization does not change it.
  • Merging sections. This cuts a path-keyed attacker from 0.81 to 0.16, but a numbering-invariant attacker (section-size profile plus cross-reference offsets) still links about 0.6.
  • Scope. You were right that this matters. Our privacy claim now covers only the text, which contains no party names or confidential details; the section tree is a quasi-identifier of any topology-preserving release, and the paper says so.

Q1: edge survival.

  • Raw source graphs have 98 (CUAD) and 151 (EDGAR) internal cross-reference edges. The k=1 targets keep 81 and 140, so the losses are preprocessing (deduplication and the one-per-clause cap).
  • 28 of 69 CUAD and 14 of 47 EDGAR documents have no internal cross-references after extraction.
  • In the headline runs the cap changes the scoring target, not only the brief. The k=4 BRIEF-L runs score against all retained edges.
  • We report raw-source F1 next to target F1. At k=4, raw-source F1 is CUAD 0.75 / 0.68 / 0.68 and EDGAR 0.58 / 0.50 / 0.48 (BRIEF / BRIEF-L / naive-L).

Q2: Eq. (1).

  • Eq. (1) is the scorer for all tables, and C includes invented edges, so they lower precision.
  • The only exclusion is a family with an empty target. The Appendix I sentence was wrong and is fixed, with pseudocode and an example: T = {1→3}, C = {1→3, 2→4} gives F1 = 0.67.

Dear Reviewer mYcS ,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

Thank you for the additional experiments and clarifications. The matched evaluation and clarification of the F1 calculation address several of my earlier questions. I also acknowledge the reduced repetition and larger coherence margin between actual and random citation targets.

However, I remain unconvinced that the results establish sufficient added utility beyond simpler baselines. The deterministic template already achieves high structural fidelity, and the citation lookup experiment preserves model rankings even under the template condition. Although the diversity improvements are useful, their downstream benefit remains insufficiently demonstrated.

Moreover, your graph matching attack still links approximately 71% to 76% of releases at k=4k=4, despite pseudonymization. The reduction in BGE linkage is meaningful for that particular attacker, but restricting the privacy claim to text does not resolve the source linkability of the complete released document.

The response strengthens the empirical analysis and appropriately narrows the claims, but my central concerns about the significance of the contribution and its practical utility remain. I therefore maintain my overall rating.

Thank you for engaging with the new results. The template matches the structure, but its releases are boilerplate: distinct-2 is 0.16 / 0.12 (CUAD / EDGAR), against 0.64 / 0.46 for the sources and 0.71 / 0.58 for BRIEF-L, and 13–15% of its clause bodies repeat. BRIEF-L is the only arm we tested that combines source-like variety, references that are coherent above a random-target control, and chance-level text linkage at bounded context. We agree that the downstream benefit of that variety is not yet measured beyond the lookup task. On graph matching: as you note, it links about 0.71–0.76 of releases, and it does so equally for the no-LLM template. It is a property of preserving the section tree, not of how the text is generated. The revised paper states this as a limit.

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

BRIEF studies how to regenerate long legal and financial documents while retaining a checkable reference topology under bounded prompts. It extracts and canonicalizes a clause graph, builds a target graph, and generates each clause from a local structural brief without source prose or sibling clauses; BRIEF-L adds dependency ordering and bounded summaries of cited clauses. Experiments cover a 70–1200-clause synthetic sweep, CUAD contracts, EDGAR EX-10 exhibits, two generators, and retrieval and language-model evaluators. BRIEF keeps maximum prompt length near 86 tokens and improves target-relative cross-reference fidelity over one-pass generation, while reducing BGE source linkage. Yet it does not consistently beat naive bounded generation on topology, and BRIEF-L changes the citation budget. Coherence improves, but semantic equivalence and formal anonymity remain unestablished.

Critical strengths
  1. The paper defines a concrete, measurable long-document regeneration task and evaluates it over a meaningful length range. In the 70–1200-clause sweep, BRIEF keeps the maximum per-call prompt near 86 tokens, section-tree fidelity at 1.0, and cross-reference fidelity near 0.98, while one-pass context grows beyond 20,000 tokens. The real-corpus comparisons and the explicit fivefold input-token overhead at 1200 clauses make the reliability–cost trade-off visible, even though the comparison to naive bounded generation is less favorable.

  2. The paper clearly exposes a distinction between structural validity and reference meaning. BRIEF’s within-release coherence is only 0.02 on CUAD and 0.06 on EDGAR despite high structural scores, while bounded cited-clause summaries raise coherence to 0.44–0.70 in the reported BRIEF-L settings. The role-agreement probe and the explicit refusal to claim semantic equivalence or formal anonymity are valuable limitations.

Critical weaknesses
  1. The central incremental contribution is not isolated by a fair comparator. BRIEF deterministically prepends canonical number paths, whereas the one-pass arm must generate numbered output; the format-matched control changes instructions but does not equalize this assembly step. Naive bounded generation matches or exceeds BRIEF on several topology results (e.g., 0.95 vs. 0.92 on Qwen CUAD and 0.92 vs. 0.82 on Llama CUAD), and naive-L has higher coherence in every Table 23 cell, albeit with higher linkage. A canonically assembled one-pass body-list baseline and a graph-conditioned rolling-summary baseline at matched token budgets are needed to show an advantage beyond scheduling and prompt enrichment.

  2. The claimed privacy–coherence trade-off is not demonstrated under a matched preserved-information channel. The source-available attacker never directly uses the canonical section/cross-reference graph that BRIEF intentionally retains, and the equal-weight text/layout fusion results do not show that layout contributes no additional identifying signal. In addition, the headline BRIEF privacy result (CUAD BGE linkage 0.37) and the BRIEF-L coherence result come from different citation setups: the summary-free BRIEF in Appendix S has linkage 0.89, while BRIEF-L has 0.88. Without a canonical-graph attack and citation-matched, source-clustered uncertainty, the added value beyond simple bounded generation remains unproven, and no downstream task shows that topology retention alone yields a useful synthetic corpus.

Questions for authors
  1. With the same deterministic numbering assembly and matched resource budgets, what advantage remains over one-pass body-list generation and a graph-conditioned rolling-summary baseline? Please provide scorer pseudocode and a worked example containing an extra candidate edge to resolve Equation (1) versus Appendix I, and specify whether Table 13 evaluates candidates or releases that passed the Appendix F acceptance checks.

  2. After canonicalizing every candidate source in the same way as the target construction, how accurately can a graph-matching attacker identify each release? Please report structure, coherence, source linkage, and total cost for BRIEF, BRIEF-L, and naive-L at the same citation budget, with source-clustered uncertainty, so the claimed trade-off can be evaluated directly.

4 comments
AuthorsOpenReview ↗

Response to Reviewer

Thank you for the precise requests. We ran the comparators you asked for, with one difference noted under W1. Setup and shared points are in the general response.

W1 / Q1: advantage over one-pass and rolling-summary under the same numbering assembly. We added three arms that use BRIEF's code-side numbering:

CUAD, k=4 (same target, same numbering assembly)target xref F1BGE linkagecoherence (random-target)tokens/doc
BRIEF0.910.890.12 (0.07)6.9k
BRIEF-L0.880.910.62 (0.14)8.8k
one-pass body list0.820.960.18 (0.12)3.2k
rolling window (brief + previous 4 clauses)0.880.900.21 (0.16)13.8k
template renderer (no LLM)0.970.890.57 (0.14)0
BRIEF-L + term pseudonymization0.850.0290.63 (0.17)8.9k
naive-L + term pseudonymization0.860.0290.64 (0.30)7.4k
  • One-pass. With numbering equalized, the per-clause schedule still improves cross-reference F1 on CUAD (0.71 vs 0.80 at k=1; 0.82 vs 0.91 at k=4) and, without pseudonyms, lowers linkage (0.74 vs 0.41). On EDGAR the two tie on F1 (0.92), and one-pass leaks more (0.57 vs 0.48); with pseudonyms both are at chance. One-pass is cheaper in tokens.
  • Rolling summary. We ran a sliding window (the brief plus the previous 4 clauses), not a graph-conditioned summary, and its token budget is not matched (13.8k vs 6.9k for BRIEF and 8.8k for BRIEF-L, which favours the window). It reaches similar F1 at about 2× tokens. Its coherence is close to its random-target control, so an unconditioned window does not produce meaningful references. Graph-conditioned cited-clause summaries (BRIEF-L) do: 0.62 vs 0.21 on CUAD, and 0.71 vs 0.22 on EDGAR.
  • Template renderer. Structure needs no generation, so we now present section-tree F1 = 1.0 as an invariant of assembly.

Naive vs BRIEF. We agree that BRIEF's prompt does not win on topology. Without pseudonyms, BRIEF-L links less than naive-L at matched k=1 (CUAD 0.52 vs 0.65; EDGAR 0.28 vs 0.36), but with term pseudonymization both reach chance (CUAD 0.029 vs 0.029; EDGAR 0.007 vs 0.000), so we no longer claim a linkage advantage. Over naive-L, whose raw coherence is equal or higher (CUAD 0.63 vs 0.64; EDGAR 0.63 vs 0.71), BRIEF-L keeps a wider real-vs-random coherence margin (0.46 vs 0.34; 0.49 vs 0.43) and, above all, does not repeat itself (0.1–0.3% of clause bodies repeated verbatim vs 12.5–14% for naive-L).

Q1: Eq. (1) vs Appendix I.

  • Every table uses Eq. (1), F1 = 2|C∩T| / (|C|+|T|), so candidate-only edges lower precision.
  • A family is skipped only when the target has no edges of that type. The Appendix I sentence was wrong and is corrected.
  • Worked example (added with pseudocode): T = {1→3}, C = {1→3, 2→4} gives P = 0.5, R = 1, F1 = 0.67.
  • One-pass loses through precision: CUAD 0.42 vs BRIEF 0.66; EDGAR 0.24 vs 0.42.
  • Table 13 scores all generated candidates, with no Appendix F acceptance filter. The caption now says so.

W2 / Q2: graph-matching attacker and matched operating point.

  • Attacker. We canonicalize every candidate source exactly as in target construction and match reference graphs (Jaccard over cross-reference edges, with a numbering tie-break). It links 0.48 (CUAD) and 0.74 (EDGAR) at the headline setting, and about 0.71–0.76 for every arm, including the no-LLM template.
  • Merging sections. This cuts a path-keyed attacker from 0.81 to 0.16, but a numbering-invariant attacker still links about 0.6.
  • Scope. The section tree is a quasi-identifier, and the template shows that any release keeping it is linkable. Our privacy claim now covers only what the text reveals: no party names or confidential details.
  • Matched point. The 0.37→0.89 linkage gap comes from defined-term names in the brief, not cross-references (see the general response). With term pseudonymization, the full retained graph and the same citation budget (k=4), BRIEF-L reaches:
    • CUAD: linkage 0.029 (chance 0.014), coherence 0.63 (random-target 0.17), target xref F1 0.85.
    • EDGAR: linkage 0.007 (chance 0.021), coherence 0.63 (0.14), F1 0.86.
    • Cost: 8.9k (CUAD) and 13.3k (EDGAR) tokens per document, against 6.9k and 10.5k for BRIEF.

Downstream use.

  • Three LLMs answered "which clause does X cite" questions on sources and releases. Their ranking is the same on every release as on the source: Llama-8B > Qwen3-4B > Qwen3-8B.
  • Accuracy is higher on releases (0.45–0.77 vs 0.29–0.44) because references are explicit.
  • We claim only that the ranking of systems on this task is preserved, not absolute accuracy.

Dear Reviewer jRGX ,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

Thank you for the additional experiments. The F1 and Table 13 issues are now clarified. However, on CUAD, pseudonymized naive-L achieves comparable or slightly better results at lower cost, so BRIEF-L’s added value remains unclear. Graph matching also still identifies sources at substantial rates, leaving the structural leakage concern unresolved. I therefore maintain my rating.

Thank you for confirming the F1 and Table 13 issues are resolved. On naive-L: with pseudonymization it matches BRIEF-L on F1 and raw coherence, but it repeats 12.5% (CUAD) and 14.4% (EDGAR) of clause bodies verbatim, against 0.1–0.3% for BRIEF-L, and its random-target coherence is higher (0.30 vs. 0.17), so its real-vs-random margin is narrower (0.34 vs. 0.46). The extra 1.5k tokens per document is what buys non-repetitive releases with references that point to the right clause. On graph matching: the attacker links about 0.72 of releases for every arm, including a no-LLM template, so this linkability comes from preserving the section tree, which any topology-preserving release must do. The revised paper states it as a limit, and scopes the privacy claim to release text.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment

General response to all reviewers

We thank the reviewers for detailed reviews. The shared concerns: (i) section-tree fidelity is built in by code-side numbering; (ii) the low-linkage headline and the coherent BRIEF-L result come from different operating points; (iii) there is no structure-aware attacker; (iv) Eq. (1) and Appendix I are inconsistent; (v) there is no downstream evidence.

New runs use Qwen3-4B, 3 seeds and document-clustered 95% CIs, with key ones replicated on Llama-3.1-8B; our setup reproduces the paper (BRIEF k=4 CUAD linkage 0.886 vs 0.89). Changes are in orange in the revised PDF.

1. Corrections.

  • Eq. (1) is the scorer used for every table. Invented edges count and lower precision. The Appendix I sentence was wrong and is fixed, with pseudocode and a worked example.
  • The one-pass arm receives no source prose (the Introduction was wrong), and its Table 7 section-tree 0.0 is an extractor convention.
  • Section-tree F1 = 1.0 is an invariant of code-side assembly, not a model capability.

2. What drives linkage. We reran BRIEF varying only the brief contents (CUAD / EDGAR BGE top-1):

brief carriesCUADEDGAR
1 xref, 1 term-use0.420.21
4 xrefs, 1 term-use0.430.20
1 xref, 4 term-uses0.890.50

The 0.37→0.89 gap comes from defined-term names, not cross-references: 47–73% of defined terms appear verbatim in releases, under 2.5% of 4-grams do.

3. A single operating point with topology, coherence and low linkage. We add defined-term pseudonymization: each defined term is replaced with a generic name from a per-document pool before generation. BRIEF-L at k=4, full retained graph:

BGE linkage (chance)coherence (random-target control)target xref F1
CUAD0.029 (0.014)0.63 (0.17)0.85
EDGAR0.007 (0.021)0.63 (0.14)0.86

On Llama-3.1-8B: linkage 0.024 / 0.014, coherence 0.50 / 0.58, F1 0.79 / 0.82.

Pseudonymization is not specific to BRIEF. Applied to the other arms at k=4 (BGE linkage / coherence with random-target control / clause bodies repeated verbatim):

+ pseudonymizationCUADEDGAR
BRIEF-L0.029 / 0.63 (0.17) / 0.1%0.007 / 0.63 (0.14) / 0.3%
naive-L0.029 / 0.64 (0.30) / 12.5%0.000 / 0.71 (0.28) / 14.4%
one-pass0.005 / 0.30 (0.25) / 2.3%0.014 / 0.25 (0.19) / 3.2%

Text-level unlinkability therefore comes from pseudonymization, for any arm. Over naive-L, BRIEF-L adds non-repetitive prose (naive-L repeats one clause body in eight) and a wider real-vs-random coherence margin (0.46 vs 0.34 CUAD; 0.49 vs 0.43 EDGAR); over one-pass, coherent references (0.63 vs 0.30).

4. Structure-aware attacker, and the claim's scope.

  • A graph-matching attacker (canonicalized sources, reference-graph matching) links about 0.72 of releases for every arm, including a no-LLM template, with or without pseudonyms. After merging sections, a numbering-invariant attacker still links about 0.6.
  • We scope the privacy claim to text: release text contains no party names or confidential details. The section tree, which we preserve by design, is a quasi-identifier of any topology-preserving release.

5. New baselines.

  • A deterministic template renderer has the best structure (F1 0.97–0.99) and the same linkage as BRIEF, but its releases are boilerplate: distinct-2 0.16 / 0.12 (CUAD / EDGAR) vs 0.71 / 0.58 for BRIEF-L and 0.64 / 0.46 for sources, with 13–15% of clause bodies repeated. Generated prose adds source-like variety, not structure.
  • One-pass with the same numbering is worse on CUAD (target xref F1 0.71 vs 0.80 at k=1) and, without pseudonyms, leaks more (0.74 vs 0.41).
  • A rolling window (brief + previous 4 clauses) has similar F1 but near-random coherence (0.21) at about 2× tokens.
  • naive-L repeats 12.5–14% of clause bodies verbatim, BRIEF-L 0.1–0.3%. Without pseudonyms BRIEF-L also links less at matched k=1 (CUAD 0.52 vs 0.65), but with pseudonyms both reach chance, so we no longer claim a linkage gain over naive.

6. Downstream task. Three LLMs answered "which clause does clause X cite?" on sources and releases. Their ranking is the same on the source and every release (accuracy is higher on releases, where references are explicit), so releases support comparing systems on this task, not absolute accuracy.

7. Coherence probe. A random-target control separates real from random cited clauses (0.44–0.70 vs 0.07–0.17); Qwen3-14B and Phi-4 agree on arm rankings (Spearman 0.94–1.0). The template also scores 0.55, so the probe measures non-contradiction more than specific fit. We will add a human check of the probe in the camera-ready.

The revised paper claims two things: BRIEF-L's cited-clause summaries give coherent references and non-repetitive prose at bounded context, and term pseudonymization brings text linkage to chance for any arm. It states structural linkability as a limit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.