{"id":"809fd80e-af91-4647-a316-fe7856bb8b6a","arxiv_id":"2505.12572","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper reports that a two-stage outline with a 0.01 compression ratio, using a mixed direct-expansion step, preserves the most semantic detail in reconstructed ultra-long novels.","lead":"This paper tests how much meaning is lost when a million-word novel is compressed into a short outline and then expanded back by an AI, and claims an outline about 1% of the original length works best. It is relevant to anyone building AI systems that write or reconstruct very long texts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the claimed optimum: K2-7 (R=0.020) has higher SemSim and StyleSim than K2-* (R=0.010), and the 'optimal' claim is only asserted under an unexplained R≤0.01 filter.","rationale":"The reader's verdict is REJECT, and I agree that the paper should be rejected in its current form. However, the single most load-bearing problem is not the sampling proxy that the reader identifies as the weakest assumption, but the direct contradiction between the paper's central claim and its own Table 2: configurations with R=0.015–0.020 achieve higher semantic and style similarity than the claimed optimum at R=0.010. The paper only ever compares K2-* against settings with R≤0.01, and the boldface in Table 2 is explicitly limited to that subset, so the claim of optimality is an artifact of an unexplained filter rather than a conclusion from the full comparison. The reader's rationale does mention the arbitrary restriction to R≤0.01, so there is partial agreement, but the reader's stated weakest assumption is the sampling validation, which is secondary. A single concrete re-analysis—ranking all configurations without the filter and testing K2-* against K2-7/K2-9—would settle whether the headline claim survives. The qualitative observation that larger outlines reduce distortion is plausible, but the paper's quantitative central claim, as stated, is not supported by the presented evidence.","tokens_in":12250,"tokens_out":5378,"duration_ms":50187,"concrete_test":"Re-rank all configurations in Table 2 without the R≤0.01 filter, using the composite distortion D defined in Section 2.4, and run a paired test (bootstrap or paired t-test) of K2-* versus K2-7 and K2-9 on SemSim, StyleSim, and the composite D. If K2-7 or K2-9 is not significantly worse on any of these metrics, the claim 'R=0.01 is optimal' is unsupported by the paper's own data; if K2-* is significantly better, the paper must still state the rate-distortion cost function under which the comparison is made.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim in Section 1.2, 'R=0.01 is the optimal compression-expansion ratio under our configuration and experiment,' is contradicted by the paper's own Table 2. K2-* (R=0.010) achieves SemSim 0.613±0.216 and StyleSim 0.611±0.133, while K2-7 (R=0.020) achieves SemSim 0.710±0.166 and StyleSim 0.654±0.110; K2-6, K2-8, and K2-9 (R=0.015–0.020) also have higher semantic and style similarity than K2-*. The table's boldface is explicitly restricted to 'tested configurations under R≤0.01,' and the significance claim in Section 4.1 is only that 'K2-* outperforms all settings with R≤0.01 (p<0.001).' No pairwise test is reported against the R>0.01 configurations, so the data as presented show distortion decreasing from R=0.001 to R=0.020. Without a stated rate cost, such as minimizing R subject to D≤ε or minimizing D+λR, the optimum is at the tested boundary R=0.020, not at R=0.010. The Limitations section explicitly declines to fix ε and excludes the intermediate outline cost from the objective, so no cost function justifies the R≤0.01 restriction. This is an internal inconsistency with the headline claim, not merely a disagreement with external conventions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies how the compression-expansion ratio R affects information distortion when reconstructing ultra-long Chinese web novels with Gemini 2.0 Flash. It compares a one-stage pipeline (novel to global outline to novel), a hierarchical two-stage pipeline (novel to section outline to global outline to novel), and a mixed two-stage pipeline (novel to section outline to global outline to novel, with the final expansion done directly from the global outline). The paper's central claims are that R=0.01 is the optimal compression-expansion ratio and that the mixed two-stage configuration K2-* (α1=0.05, α2=0.20) yields the best fidelity. The evidence consists of eight sampled chapters per novel across 40 novels, scored by GPT-4o on semantic, plot, character, background, and style similarity, plus structural difference metrics.","tokens_in":12554,"tokens_out":3811,"duration_ms":37753,"significance":"If the central claim were correct, the paper would provide a practically useful design rule for hierarchical long-form novel generation and an interesting empirical connection between rate-distortion ideas and LLM-based summarization and expansion. The paper has several strengths: it uses a substantial corpus of 40 million-word novels, evaluates with multiple complementary metrics, reports pairwise significance tests, and states its limitations candidly. However, the headline optimality claim is contradicted by the paper's own Table 2, and the sampling justification rests on an undescribed pilot study. As it stands, the paper does not establish the claimed optimum, so its main contribution is not supported.","major_comments":[{"comment":"The claim in Section 1.2 that 'R=0.01 is the optimal compression-expansion ratio under our configuration and experiment' is contradicted by Table 2: K2-* (R=0.010) has SemSim 0.613±0.216 and StyleSim 0.611±0.133, whereas K2-7 (R=0.020) has SemSim 0.710±0.166 and StyleSim 0.654±0.110, and K2-6, K2-8, and K2-9 (R=0.015-0.020) also exceed K2-* on these metrics. The table's boldface is explicitly restricted to 'tested configurations under R≤0.01', and the significance test in Section 4.1 compares only settings with R≤0.01; no pairwise test is reported against R>0.01 configurations. Without a stated cost constraint such as minimizing R subject to D≤ε or minimizing D+λR, the observed optimum lies at the tested boundary R=0.020, not R=0.010. The Limitations section explicitly declines to fix ε and excludes the intermediate outline cost, so the reported data do not support the headline optimality claim.","section":"§1.2 and Table 2"},{"comment":"The claim that sampled-chapter distortion approximates full-book distortion rests entirely on a pilot study over 40 books with r=0.95, but no details of the pilot, its protocol, or how full-book distortion was computed are provided. The Fisher z-transform sentence is asserted without the corresponding analysis. Since every reported distortion value is computed from only eight sampled chapters per novel, an unrepresentative pilot correlation would invalidate the measured optimal ratio for whole novels; this is load-bearing for the paper's central claim and needs to be documented or the sampling claim must be softened accordingly.","section":"§4.1 Sampling Justification"},{"comment":"The rate measure R=α1×α2 omits the token cost of the intermediate section outline, which is α1×L0. Consequently, comparisons between K=1 and K=2 configurations at the same nominal R do not compare equal total compression effort, and the 'optimal ratio' is not an optimal rate in the sense of the stated rate-distortion objective. The Limitations paragraph acknowledges that the intermediate cost is excluded but does not quantify the effect or provide a sensitivity analysis; for a paper whose title names an optimal expansion ratio, the excluded cost should enter the objective or be shown not to change the ranking.","section":"§2.1 and Limitations"},{"comment":"The paper's notion of 'optimal' is circular in practice: R is selected as the argmax of the measured similarity curve over a small grid, so the conclusion restates the best grid point rather than providing an independent prediction. Section 6 and Appendix B report only a weak correlation between R and similarity (r=0.129, p<0.001), which further undermines the idea that a sharp optimum at R=0.01 has been identified. The authors should either formulate a testable rate-distortion trade-off or explicitly reframe the contribution as a comparison of a small set of configurations.","section":"§4.1, §6, and Appendix B"}],"minor_comments":[{"comment":"Please use 'ultra-long' consistently and fix the verb form in the abstract ('significantly reduce' should be 'significantly reduces').","section":"Abstract and title"},{"comment":"The dataset description says 10 novels were chosen from each of four genres, while Appendix J describes PPS sampling with Neyman allocation; please reconcile these two descriptions or clarify how the 40 novels were actually selected.","section":"Section 2.2 vs. Appendix J"},{"comment":"The sentence 'We implement a two-stage sampling design :' is duplicated verbatim at the start of Appendix J and should be removed once.","section":"Appendix J"},{"comment":"The prompt template in Appendix F states a 200-300 word requirement for the plot summary introduction, but the accompanying JSON schema says '400-500 words'; please make these numbers consistent.","section":"Appendix F"},{"comment":"There is a missing space in 'Bconfirms' in the first paragraph of Section 4.1.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The self-contained contradiction between the abstract/Section 1.2 and Table 2 is decisive: the paper's own numbers show that the best tested configuration is not K2-* at R=0.010. Because the central claim is the paper's main contribution, and the manuscript currently provides no cost model that would justify restricting attention to R≤0.01, I do not see how a modest revision could repair the headline result without substantial new analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious empirical study of a real niche—hierarchical outline compression and expansion at the million-word scale—and the qualitative findings are plausible. But the headline number, R=0.01 as optimal, is not supported by the paper's own data. Table 2 shows K2-7 at R=0.02 with SemSim 0.710 and StyleSim 0.654, beating K2-* at R=0.01 with 0.613 and 0.611. The boldface is restricted to R≤0.01, and the significance test is only against other R≤0.01 settings. Without a cost function that penalizes outline tokens, the data show distortion generally decreasing from R=0.001 to R=0.020, so the optimum sits at the tested boundary, not at 0.01. The Limitations explicitly decline to fix epsilon and exclude intermediate outline cost, so the restriction is arbitrary.\n\nWhat is genuinely new is the scale. Prior work mostly goes to 100k words; here they compress and expand 1M-word Chinese web novels, with a stratified sampling design, detailed prompts, and LLM-judge metrics. The observation that a two-stage outline helps, and that direct expansion from the global outline beats reinstating the intermediate section outline, is a useful engineering result. The appendix material is candid about several limitations.\n\nSoft spots: the information-theoretic framing adds little—R is defined as a product of ratios, then the optimum is read off a grid. The sampling proxy (8 chapters per novel) is justified only by a pilot correlation r=0.95 with no details. Pairwise significance is reported, but the key comparison against R>0.01 is missing. LLM-as-judge bias is acknowledged but not corrected.\n\nThis paper is for people building long-form generation pipelines and outline-to-manuscript systems. They will find the prompt engineering and scale useful even if they ignore the optimality claim. It deserves referee time because the empirical setup is substantial and the flaw is fixable: re-define the objective, test against all grid points, and report the sampling validation. As it stands, reject, but with a clear path to revision.","headline":"Useful million-word-scale study, but the R=0.01 optimality claim is contradicted by the paper's own Table 2 and needs a cost function and full-grid significance testing.","tokens_in":13129,"tokens_out":1743,"would_cite":false,"duration_ms":18155,"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":"A 1% compression-expansion ratio best preserves meaning when rebuilding million-word novels, the paper argues.","keywords":["ultra-long novel generation","hierarchical outline pipeline","compression-expansion ratio","rate-distortion theory","semantic distortion","LLM-based reconstruction","Chinese web novels","structured JSON outlines"],"falsifier":"Score every chapter—not just eight per novel—of a set of completely reconstructed million-word novels and re-estimate distortion for the tested ratios $R=0.001, 0.005, 0.01, 0.015, 0.02$ and configurations. If the minimum distortion no longer falls at $R=0.01$ with $\\alpha_1=0.05$, $\\alpha_2=0.20$, and direct expansion, then the central optimal-ratio claim fails.","tokens_in":11981,"feed_emoji":"📚","tokens_out":12528,"duration_ms":106453,"temperature":0.7,"pith_summary":"The paper asks how much of a million-word novel's meaning survives being compressed into an outline and then expanded back by a large language model. It argues that the compression-expansion ratio—outline length divided by novel length—is the quantity that controls semantic and structural distortion, and that under its tested configuration the optimum is $R=0.01$: a 10,000-word outline for a 1,000,000-word novel. It further claims that the best pipeline is a mixed two-stage design with $\\alpha_1=0.05$ (about 50,000 words) followed by $\\alpha_2=0.20$ (about 10,000 words) and direct expansion from the global outline to chapter text, using structured JSON outlines. If correct, ultra-long novel generation can be steered by outlines as small as 1% of the original, and spending more than about 1% on the final outline buys little fidelity. The paper also finds that fidelity gains from enlarging the outline are steep from 0.1% to 1% but nearly flat from 1% to 2%.","feed_headline":"A 1% outline best rebuilds million-word novels","feed_subtitle":"Two-stage compression to a 10,000-word global outline, then direct expansion, preserves the most plot and characters.","key_machinery":"The load-bearing mechanism is a hierarchical two-stage compression-expansion pipeline with per-stage rates $\\alpha_1$ and $\\alpha_2$, combined with a composite distortion measure. The overall compression ratio is the product $R=\\alpha_1\\alpha_2$, and the objective is framed as a rate–distortion problem: minimize outline rate $R$ subject to reconstruction distortion below a bound $\\epsilon$. Distortion is computed as a mixture of one minus cosine similarity, one minus BERTScore, an LLM-judged score across semantic, plot, character, background, and style similarity, and absolute differences in counts of characters, scenes, and props. The structured JSON outline—listing a plot-summary lead, characters, props, scenes, and foreshadowing—is what makes hierarchical compression reversible enough that direct expansion from the global outline outperforms staged expansion.","core_discovery":"On its own terms, the paper's central claim is that under the tested setup—a specific large language model, Chinese web novels, and the paper's compression and expansion prompts—semantic and structural distortion in reconstruction is minimized at compression-expansion ratio $R=0.01$, meaning a global outline of about 10,000 words for a 1,000,000-word novel. The best-performing configuration is the mixed two-stage pipeline it calls K2-*: compress the novel first to a section-level outline at $\\alpha_1=0.05$ (roughly 50,000 words), then to a global outline at $\\alpha_2=0.20$ (roughly 10,000 words), and expand directly from the global outline to chapter text, skipping the intermediate outline. This configuration beats one-stage compression at the same ratio and beats full two-stage expansion on the paper's similarity and structural metrics. The paper also reports that distortion falls more steeply as the outline grows from 0.1% to 1% of the original than from 1% to 2%, so outline length has diminishing returns beyond the optimum.","pith_inferences":["The paper's rate term $R=\\alpha_1\\alpha_2$ counts only the final global outline, not the intermediate section outline, so the literal end-to-end compression cost of the best configuration is about 6% of the original (50,000 words of section outline plus 10,000 words of global outline), not 1%; readers applying the ratio should budget accordingly.","The K2-* result suggests a design rule worth testing in other long-form generation settings: compression should be multi-level, but expansion should be single-step; applying that rule to textbooks, screenplays, or transcripts is a natural next experiment.","If the claimed $r=0.95$ correlation between sampled and full-book distortion holds on independent books, the eight-chapter sampling scheme would cut evaluation cost by roughly 96% and could become a reusable proxy for million-word-scale reconstruction studies.","Outlines larger than about 2% of the novel may approach the model's output-token ceiling and cause truncation rather than better fidelity; the paper does not test that regime, so the shape of the curve beyond $R=0.02$ is an open question."],"forward_implications":["For the tested model and genre mix, a global outline of about 1% of the original length is enough; longer outlines add little fidelity, so practitioners can save tokens by stopping near $R=0.01$.","Compress hierarchically but expand directly: the K2-* shape beats both one-stage and full two-stage reconstruction at the same overall ratio.","Structured JSON outlines that track characters, props, scenes, and foreshadowing preserve more structure than plain narrative outlines of equal length.","Fidelity gains from outline length are strongly diminishing: the step from 0.1% to 1% helps more than doubling from 1% to 2%.","The claimed optimum is tied to the tested configuration; changing model, language, or prompts is expected to move the optimum."],"supporting_citations":[{"why":"Supplies the LLM-as-compressor (LLMZip) method the paper builds on for analyzing lossy compression with language models.","marker":"Valmeekam et al., 2023"},{"why":"Provides the rate–distortion formulation used to define the optimization objective over outline rate $R$ and distortion bound $\\epsilon$.","marker":"Blau and Michaeli, 2019"},{"why":"LongWriter serves as the one-stage baseline (method D) that the hierarchical pipeline must beat.","marker":"Bai et al., 2024"},{"why":"Describes the Gemini model family used as the reconstruction backbone.","marker":"Team et al., 2024"},{"why":"Documents the GPT-4o model used as the LLM judge producing the semantic similarity scores.","marker":"OpenAI et al., 2024"},{"why":"Supplies the 40 Chinese web novels used for all experiments.","marker":"zxbsmk, 2023"},{"why":"Neyman allocation determines per-genre sample sizes in the two-stage sampling design.","marker":"Olayiwola et al., 2013"},{"why":"Defines text entropy by word count, the practical stand-in for the rate term $H(\\cdot)$.","marker":"Bentz et al., 2017"}],"fun_headline_variants":["1% outline minimizes distortion in novel reconstruction","Optimal outline: 1% of novel length, then expand directly","Million-word novels rebuild best from 1% global outline","Diminishing returns: novel outlines beyond 1% lose fidelity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion assumes that distortion measured on eight sampled chapters per novel accurately represents whole-novel distortion, even though each million-word novel has roughly two hundred chapters.","fun_headline_variants_meta":{"raw":{"variants":["1% outline minimizes distortion in novel reconstruction","Optimal outline: 1% of novel length, then expand directly","Million-word novels rebuild best from 1% global outline","Diminishing returns: novel outlines beyond 1% lose fidelity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000292,"raw_usage":{"total_tokens":1679,"prompt_tokens":894,"completion_tokens":785,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":715}},"tokens_in":510,"tokens_out":785,"duration_ms":7849,"temperature":1.0,"reasoning_tokens":715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:30:55.691557+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Score every chapter—not just eight per novel—of a set of completely reconstructed million-word novels and re-estimate distortion for the tested ratios $R=0.001, 0.005, 0.01, 0.015, 0.02$ and configurations. If the minimum distortion no longer falls at $R=0.01$ with $\\alpha_1=0.05$, $\\alpha_2=0.20$, and direct expansion, then the central optimal-ratio claim fails.","supporting_citations":[],"review_version":1}