{"id":"1471d506-a499-41ef-a304-17295fa06ed5","arxiv_id":"2505.07888","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-layered sentence and paragraph template method for zero-shot long-text style transfer, with a reported average gain of 0.20 over direct prompting but limited statistical and external support.","lead":"This paper proposes ZeroStylus, a zero-shot framework that rewrites long texts in another author's style by extracting separate sentence and paragraph templates from reference papers. It reports small average gains over direct prompting, but the style metric is partly tied to the model that creates the templates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Style-consistency score uses the same encoder that conditions generation, so the 6.90 vs 6.70 gap may measure template proximity rather than transferred style.","rationale":"The paper proposes a coherent two-phase framework, and the adversarial ablation comparing SentencePattern and StructuredRewritten (Table 3) provides some independent evidence for paragraph-level encoding improving content preservation: the content-preservation win rate favors StructuredRewritten (57% vs 43%), and that dimension uses BLEURT and keyword retention rather than πenc. I therefore do not claim the method is vacuous. However, the strongest claim in the abstract and Table 1, the 6.90 vs 6.70 average, is not supported by the current evidence because the dominant component x is computed with the same encoder that selects and refines templates. This is exactly the reader's weakest assumption, and the paper's own limitation discussion about missing long-text style metrics reinforces it. A conditional verdict is appropriate: the framework is plausible, but acceptance should require external scoring of style consistency and ideally artifact release. Since the reader already assigned CONDITIONAL, my stress-test does not move the verdict.","tokens_in":10811,"tokens_out":5061,"duration_ms":52684,"concrete_test":"Recompute Table 1's X column and the 6.90/6.70 averages using an external style signal: train a style or author classifier on held-out target-author papers with a different embedding backbone than πenc (for example, a fine-tuned RoBERTa or a BERT classifier), use its confidence as the style-consistency score, renormalize with the same min-max procedure, and recompute the averages over the N=500 samples. If the X advantage (7.39 vs 6.42) shrinks below the Y disadvantage or fails a significance test, the central claim is an artifact of shared πenc; if the gap survives external scoring, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest empirical claim, StructuredRewritten's 6.90 average tri-axial score versus DirectPrompt's 6.70, rests primarily on the style-consistency component x, which is computed as paragraph-level embedding similarity between output and reference texts using the encoder πenc. The same encoder drives generation: Phase 2.1 (Eq. 3) selects the paragraph template τ*_p by minimizing ||esrc_p − τp|| in πenc space, and Phase 2.3 refines the output to align with τ*_p. A method explicitly conditioned to move source embeddings toward centroids of reference embeddings will tend to score higher on a similarity metric computed with that same encoder, regardless of whether human readers perceive the target style. Table 1 confirms the advantage over DirectPrompt is concentrated in X (7.39 vs 6.42), while Y (7.04 vs 7.34) and Z (6.26 vs 6.34) are actually lower; the average gain is thus driven by the potentially circular component. The paper offers no calibration of πenc against human style judgments, and the Discussions concede that dedicated long-text style benchmarks are missing. This does not prove the method fails; it means the headline improvement is currently unverified as evidence for the dual-layer mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ZeroStylus, a zero-shot, training-free framework for long-text style transfer. It builds two hierarchical template repositories from reference texts—sentence-level patterns and paragraph-level structural templates—and uses them to guide an LLM through template-matched sentence rewriting and paragraph-level coherence refinement. The authors evaluate on academic-paper style transfer using 500 samples and report an average tri-axial score of 6.90 for StructuredRewritten versus 6.70 for DirectPrompt, plus adversarial win-rate comparisons among ablations. The central claim is that the dual-layer template mechanism preserves paragraph-level semantic coherence while retaining sentence-level stylistic strength, outperforming direct prompting and sentence-only variants.","tokens_in":11082,"tokens_out":5356,"duration_ms":56761,"significance":"If the empirical claims held, the paper would make a useful contribution: it addresses a real gap—style drift in long-text transfer—and does so with an interpretable, modular pipeline that avoids parallel corpora and fine-tuning. The method is transparently specified, uses public datasets, and includes ablation-style comparisons against several baseline categories. The authors also candidly acknowledge the lack of dedicated long-text style benchmarks. However, the current evaluation does not establish the headline improvement: the style-consistency metric is computed with the same encoder that drives template retrieval, the reported gain is concentrated in that metric, and key ablation tables are internally inconsistent. These issues affect the central empirical assertion, so the manuscript is not yet acceptable in its present form.","major_comments":[{"comment":"The style-consistency axis x is defined as paragraph-level embedding similarity between output and reference texts computed via πenc, and the same πenc is used in Phase 2.1 to select sentence and paragraph templates (Eqs. (2)-(3)) and in Phase 2.3 to align the output with the retrieved paragraph template. A method explicitly conditioned to move source embeddings toward reference-template centroids will therefore score higher on x regardless of whether human readers perceive the target style. The observed overall advantage of StructuredRewritten over DirectPrompt (6.90 vs 6.70) is driven entirely by x (7.39 vs 6.42), while y (7.04 vs 7.34) and z (6.26 vs 6.34) are lower. The paper offers no validation that πenc similarities correspond to human judgments of style. Please provide an independent style-consistency assessment (e.g., human ratings or a held-out encoder/classifier not used in generation) and report confidence intervals or significance tests for all axes.","section":"Experiments: Benchmarking Style Transfer Quality; Methods, Eqs. (2)-(3)"},{"comment":"The adversarial evidence for paragraph-level structural encoding is internally inconsistent. Table 3's caption says 'between SentencePattern and StructuredRewritten,' but the table header row says 'TemplateOnly vs StructuredRewritten.' The accompanying text claims that StructuredRewritten improves semantic preservation with a '57% vs 43% win rate,' yet the Content Preservation (Y) row in Table 3 shows 46, 39, and 44, which averages to 43%, not 57%; the 57% figure actually appears in Table 2's Style Consistency row. Please correct the captions/headers, clarify which method's win rate is reported in each table, and reconcile the text with the numbers, since this paragraph is a load-bearing part of the ablation argument.","section":"Adversarial Evaluation: Table 3 and Result text"},{"comment":"The qualitative summary overstates what Table 1 shows. The text says StructuredRewritten 'preserves TemplateOnly's stylization strength,' but its Style Consistency X score is 7.39, below TemplateOnly's 7.62; it also says the method 'maintains DirectPrompt's paragraph-level semantic coherence,' but its Content Preservation Y is 7.04, below DirectPrompt's 7.34. The average gain of 0.20 therefore does not correspond to the described component-wise behavior. Either soften these claims or support them with significance testing of the per-axis differences.","section":"Experiments: Result, Table 1"},{"comment":"All adversarial win rates are based on N1 = 100 samples (Eq. (10)) with a decision margin δ = 0.1, but no confidence intervals or significance tests are reported. Win rates of 52-55%, such as the X row in Table 3 (52, 54, 55), are within the range of plausible sampling noise at this sample size, so the conclusions drawn from these numbers—e.g., that the two-stage framework 'keeps close' in stylization strength or that paragraph-level encoding improves content preservation—are not yet statistically supported. Please report intervals or a formal test, and state the effective sample size after the δ threshold is applied.","section":"Adversarial Evaluation: Eq. (10) and Tables 2-3"}],"minor_comments":[{"comment":"In Eqs. (7)-(8), the symbol s(A,B)_M is used for two different preference scores; rename them, for example s_A^M and s_B^M, to avoid confusion.","section":"Methods, Eqs. (7)-(8)"},{"comment":"Eq. (2) uses arg max over similarity while Eq. (3) uses arg min over distance; please state the sign convention (e.g., distance as negative similarity) or make both equations use the same convention.","section":"Methods, Eqs. (2)-(3)"},{"comment":"The Original and Style rows have no Average value; since the average is the headline metric, mark these as non-applicable or compute them for completeness.","section":"Experiments, Table 1"},{"comment":"The text refers to 'as shown in 1' and 'as in 2'; please use 'Figure 1' and 'Figure 2' consistently.","section":"Methods, Figures 1-2"},{"comment":"There is a typo in the Discussion section: 'onfirming' should be 'confirming'.","section":"Discussions"},{"comment":"The evaluation averages over GPT-4o and DeepSeek-R1 but reports only the mean; given that the same models serve as encoder, extractor, and generator, per-model breakdowns would help separate the method's effect from the model's effect.","section":"Experiments: Setup"}],"recommendation":"major_revision","confidential_remarks":"The circularity concern in the reader's report is confirmed by the manuscript's own equations and metric definitions; this is a substantive correctness risk for the central claim, not a mere presentation issue. The ablation-table inconsistency and the lack of significance testing compound the problem. I would be willing to consider a revised version that adds an independent style-consistency evaluation and corrects the reported tables and claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read through ZeroStylus, the long-text style transfer paper. The core idea—extract sentence- and paragraph-level templates from reference texts and use them to condition LLM rewriting—is sensible and clearly presented. That is the real contribution: a concrete two-level template repository with DBSCAN clustering and distance-based matching, addressing a genuine problem (style drift in long documents). The method is zero-shot and needs no parallel data, which is practically useful. The ablation structure is also reasonable, isolating the effect of paragraph templates.\n\nThe elephant in the room is the evaluation. The headline 6.90 vs 6.70 gain over DirectPrompt is driven almost entirely by the style consistency axis (7.39 vs 6.42). That axis is computed as embedding similarity using the same encoder (πenc) that selects the templates in Eqs. (2)-(3). The method is explicitly designed to pull output embeddings toward reference template centroids, so a similarity score computed with that same encoder will reward exactly that behavior. There is no human calibration of πenc, no independent style metric, and no significance test on the 0.20 difference. Meanwhile content preservation and expression quality actually drop slightly relative to DirectPrompt, which undercuts the average-gain interpretation. This is a load-bearing flaw: the paper claims the full pipeline outperforms direct prompting, but the evidence is a self-referential metric.\n\nOther soft spots: the 'adversarial evaluation' is really pairwise LLM-as-judge, not adversarial in any technical sense; table labels are confused (Table 2/3 mix-ups); no code, prompts, or hyperparameters are released; and the authors themselves note the lack of dedicated long-text style benchmarks. The content-preservation ablation (57% vs 43% win rate for StructuredRewritten) uses BLEURT and keyword retention, which are at least partially independent, so that finding is more credible—but it still does not establish superiority over the direct baseline.\n\nOverall, the paper is worth reading for the two-level template architecture and as a cautionary case of metric circularity. The idea deserves refinement, not burial. I would send it to peer review only if the venue insists on reproducible evidence; if I were the editor, I would ask for an independent style measure, significance testing, and artifact release before accepting it. In its current form, the load-bearing empirical claim is unverified.","headline":"A sensible two-level template pipeline for zero-shot long-text style transfer, but the headline gain over direct prompting rests on a self-referential style metric that does not establish the claimed advantage.","tokens_in":11568,"tokens_out":2902,"would_cite":false,"duration_ms":29332,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Zero-shot long-text style transfer improves when rewrites follow both sentence- and paragraph-level templates.","keywords":["text style transfer","long-text generation","zero-shot learning","template-based rewriting","paragraph structure","discourse coherence","large language models"],"falsifier":"Have independent human annotators rate style consistency on the 500 test outputs without access to $\\pi_{\\mathrm{enc}}$, and compare StructuredRewritten against DirectPrompt; if the advantage disappears, the reported 6.90 versus 6.70 gap depends on the shared encoder rather than on genuine style transfer.","tokens_in":10597,"feed_emoji":"📝","tokens_out":9510,"duration_ms":86225,"temperature":0.7,"pith_summary":"This paper argues that zero-shot style transfer of long text fails when style is treated as a sentence-level property, because paragraph-level rhetorical structure is itself part of the style. It proposes a two-phase framework that first extracts reusable sentence templates and paragraph templates from reference texts, then rewrites each source sentence under both the best-matching sentence template and the best-matching paragraph template. On a 500-sample academic-writing evaluation, the full method averages 6.90 on style consistency, content preservation, and expression quality, against 6.70 for direct prompt-based rewriting. The implication, if the result holds, is that coherent long-text style transfer is reachable without parallel corpora or LLM fine-tuning.","feed_headline":"Two-level style templates beat direct LLM prompting on long text","feed_subtitle":"Matching sentence and paragraph patterns separately keeps content and coherence in zero-shot rewriting.","key_machinery":"The load-bearing object is the pair of template repositories $\\Gamma_s$ and $\\Gamma_p$, built and used by the same encoder $\\pi_{\\mathrm{enc}}$. Sentence embeddings are clustered with a density-based algorithm (DBSCAN) into prototypical sentence templates; paragraph embeddings, formed by encoding the sequence of sentence embeddings, become paragraph templates whose addition is gated by a distance threshold. At rewrite time, Eq. (2) selects $\\tau_s^i = \\arg\\max_{\\tau \\in \\Gamma_s} \\mathrm{sim}(e_i^{\\mathrm{src}}, \\tau)$ for each sentence and Eq. (3) selects the paragraph template $\\tau_p^* = \\arg\\min_{\\tau_p \\in \\Gamma_p} \\|e_p^{\\mathrm{src}}-\\tau_p\\|$; the generator then rewrites each sentence under this triple conditioning, and a refinement step adjusts transitions and discourse markers to fit the paragraph template. This mechanism is what lets the method decouple local expression from global structure.","core_discovery":"The central claim is that effective long-text style transfer needs style modeling at two granularities at once, and that this can be done zero-shot by treating style as a set of reusable structural templates. The paper introduces ZeroStylus, whose first phase embeds sentences from reference texts, clusters them into a sentence template repository $\\Gamma_s$, and encodes whole paragraphs into a paragraph template repository $\\Gamma_p$; the second phase matches each source sentence to its nearest sentence template and the source paragraph to its nearest paragraph template, then generates and refines the rewritten paragraph under both constraints. The authors report that this dual-layer design yields an average tri-axial rating of 6.90, above 6.70 for direct prompting and 6.62 for the sentence-template-only ablation, with adversarial comparisons showing that adding paragraph-level structure mainly improves content preservation. The paper states that the full method keeps the stylization strength of template-only rewriting while retaining the semantic coherence of direct prompting.","pith_inferences":["Because the same encoder guides template retrieval and scores style consistency, the headline advantage may be inflated by shared representational biases; a human-judged style rating on the same outputs would test whether the 6.90 versus 6.70 gap is genuine style transfer or template overlap.","The decoupling of sentence and paragraph templates suggests a cheap update path for personal writing assistants: swapping in a new author's paragraph templates should restyle a document's organization while reusing generic sentence patterns, which is testable without retraining.","A natural stress test is length: if the paragraph template is doing the work, the advantage over direct prompting should grow as paragraphs get longer and richer in discourse structure; if it does not, the mechanism is mostly sentence-level after all.","The recipe may transfer to other long-form style tasks, such as dialogue persona or document-level formality, by replacing paragraph templates with turn-level or section-level structural templates."],"forward_implications":["Zero-shot long-text style transfer becomes feasible without parallel corpora or fine-tuning, since the only resources are reference texts and an LLM used at inference time.","Sentence-level transfer methods should be expected to lose inter-sentence coherence on long inputs; adding a paragraph-level template recovers much of that loss, mainly in content preservation.","The template repository is incremental: new paragraph templates enter only when their embedding is farther than a threshold, so style updates could be applied by adding a subset of new references rather than reprocessing an entire corpus.","Length-constrained, segment-by-segment rewriting is part of the method's defense against style drift, so the framework should scale to documents longer than a single LLM context window."],"supporting_citations":[{"why":"Supplies the ConvTransfer baseline, a few-shot conversation style transfer method the paper compares against as a sentence-level approach.","marker":"(Roy et al. 2023)"},{"why":"Defines non-parallel author-stylized rewriting and the DirectPrompt-style zero-shot baseline, plus the style-alignment evaluation dimensions.","marker":"(Syed et al. 2020)"},{"why":"Provides the ArxivPapers dataset from which source paragraphs and reference style texts are sampled.","marker":"(Kardas et al. 2020)"},{"why":"Provides the Arxiv 10 dataset used together with ArxivPapers for the 500 test samples.","marker":"(Farhangi et al. 2022)"},{"why":"Supplies the keyword and abstract field-alignment step that pairs each source paragraph with relevant reference articles before transfer.","marker":"(He et al. 2025)"},{"why":"Is one of the two large language models used as encoder, extractor, and transferer in the pipeline.","marker":"(OpenAI et al. 2024)"},{"why":"Is the second large language model backbone whose outputs are evaluated and averaged with the first.","marker":"(DeepSeek-AI 2025)"}],"fun_headline_variants":["Dual-tier templates beat direct prompting for long-text style","Zero-shot rewrite: sentence and paragraph templates win","Two-level style mapping improves zero-shot long-text transfer","ZeroStylus: dual templates outperform plain LLM prompting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework and its headline score rest on a single encoder: the same text-embedding model retrieves matching templates and then measures style consistency, so if that model does not actually capture paragraph-level style, both the rewriting signal and the reported score fail together.","fun_headline_variants_meta":{"raw":{"variants":["Dual-tier templates beat direct prompting for long-text style","Zero-shot rewrite: sentence and paragraph templates win","Two-level style mapping improves zero-shot long-text transfer","ZeroStylus: dual templates outperform plain LLM prompting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1546,"prompt_tokens":965,"completion_tokens":581,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":517}},"tokens_in":581,"tokens_out":581,"duration_ms":6519,"temperature":1.0,"reasoning_tokens":517,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:31:00.608343+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have independent human annotators rate style consistency on the 500 test outputs without access to $\\pi_{\\mathrm{enc}}$, and compare StructuredRewritten against DirectPrompt; if the advantage disappears, the reported 6.90 versus 6.70 gap depends on the shared encoder rather than on genuine style transfer.","supporting_citations":[{"cited_title":"Adapting Language Models for Non-Parallel Author-Stylized Rewriting","cited_arxiv_id":"1909.09962","evidence_quote":"Defines non-parallel author-stylized rewriting and the DirectPrompt-style zero-shot baseline, plus the style-alignment evaluation dimensions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Arxiv 10 dataset used together with ArxivPapers for the 500 test samples."}],"review_version":1}