MUSE: A Theory-Harnessed Story Engine for Vibe Narrativizing
Jianxiang Ma, Xiaocui Yang, Daling Wang, Yuesong Hou, Mingfu Zhang, Yichen Gao, Junzhao Huang
cs.CL
Quick verdict
McKee-theory agent engine lifts LLM story quality on 2 of 3 benchmarks across 4 models; promising, but all-LLM judging and weak baselines temper it.
Summary
The paper formulates “Vibe Narrativizing” — converting a finite natural-language writing brief into a finished story — and presents MUSE, a multi-stage agent system built on Robert McKee’s screenwriting theory. Knowledge engineering turns the theory into a rule library through rule atomization, semantic consolidation, mechanism abstraction, a single source of truth, and layered disclosure; an eight-phase harness (design, character performance, scene writing, layered review) plus a canon corpus of 31 masterworks (572 scenes) preserve story decisions in persistent artifacts across stages. On the 58-query WritingBench novel-creation subset and a 38-query ConStory-Bench sample, judged by claude-sonnet-4-6, MUSE gains 1.1–6.2 WritingBench points over zero-shot on Claude Sonnet 4.6, Opus 4.7, GPT-5.4, and GPT-5.5, and is the only multi-stage system to beat zero-shot — four reproduced baselines score 7.2–37.7 points below their base models. LongStoryEval rises by more than ten points on three of the four models. Consistency error density stays at 0.245–2.117 errors per 10k words, below reproduced baselines but above zero-shot on all four models. Ablations (four-model averages) attribute the largest effect to outline design (−2.25 WritingBench, −6.42 LSE when removed). Acknowledged limitations: no human evaluation and unreported token cost; no significance tests or run-to-run variance are reported, and baseline reproduction fidelity is unverified.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Vibe Narrativizing can be formalized as mapping a natural-language brief to a finished story plus a persistent decision trace jointly satisfying four requirements: intent reflection, beat fidelity, voice differentiability, and coherence (Eq. 2).
- Encoding McKee's story theory through five knowledge-engineering operations (atomization, consolidation, mechanism abstraction, SSOT, layered disclosure) yields guidance that four frontier LLMs can follow across planning, drafting, and revision.
- MUSE is the only multi-stage system among five compared to improve WritingBench over zero-shot on all four base models (+1.1–6.2 points, Table 1).
- MUSE raises LongStoryEval by more than ten points on three of four models while keeping consistency error density at 0.245–2.117 per 10k words, below reproduced baselines (though above zero-shot).
- Component ablations attribute the largest quality contribution to structural outline design (−2.25 WB, −6.42 LSE when removed), with character packages mainly reducing AI-detectable prose and review adding revision gains.
Strengths
Well-specified mechanism chain from theory to decisions
Beat contracts b_i = (v−, v+, p) with a charge map σ (Eq. 1), a per-artifact well-formedness predicate Φt (Eq. 10), named review failure modes (non-event, on-the-nose dialogue, static arc, removable peak — Table B.2), and role-scoped context assembly (Table C.1) make the pipeline concrete; Appendix J traces query 189 end-to-end from a rule to the climactic action.
Broad, multi-model evaluation matrix
Four frontier base models × five systems × three benchmarks (Table 1), plus per-domain WritingBench breakdown (Figure 3: MUSE leads zero-shot on 6/7 domains under GPT-5.5, +13.0 Fan Fiction, +11.8 Derivative Work) and per-model, per-domain ablations (Appendix I). WB gains of 1.1–6.2 reproduce across very different base models.
Self-instrumentation and candid ablations
Coherent ablation narrative matching the three-path theory
Reproducibility scaffolding
Weaknesses
Unvalidated baseline reproductions carry the headline claim
Consistency regression relative to zero-shot is obscured by framing
Judge-only evaluation with a same-family judge
No error bars, no cost accounting, no canon ablation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MUSE (arXiv 2609.15188v3, Sept 2026, cs.CL) formulates **Vibe Narrativizing** — mapping a finite natural-language brief to a finished story plus an ordered artifact trace (Eqs. 1–3) — and attacks two named bottlenecks: *rule quality* (whether story knowledge yields clear guidance) and *rule realization* (whether decisions persist through generation). The response is twofold: knowledge engineering of McKee's *Story* into a rule library via atomization, semantic consolidation, mechanism abstraction, single-source-of-truth, and layered disclosure (§3.2), and an eight-phase agent harness (§3.3, App. C) whose persistent artifacts — beat contracts b_i = (v−, v+, p), character packages, scene cards — carry decisions across Design (Phases 0–5), Performance, Creation, Review, and Canon Reference (31 masterworks, 572 scenes). Acceptability is formalized as four conjoint requirements (Eq. 2). Headline claims: +1.1–6.2 WritingBench (WB) over zero-shot on all four base models; the only multi-stage system in the comparison to beat zero-shot; LongStoryEval (LSE) +10 on three of four; ConStory-Bench error density (CED) 0.245–2.117 per 10k words.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is 'Vibe Narrativizing'?
It is this paper's name for the task of turning a finite natural-language writing request (genre, characters, required events, prose style) into a finished story, analogous to 'vibe coding': the user states intent in natural language and the system carries out the full production — planning artifacts, scene drafts, and revisions — not just a single completion.
Does MUSE fine-tune the underlying language models?
No. MUSE is a prompt-and-agent harness on top of frozen frontier LLMs (Claude Sonnet 4.6, Opus 4.7, GPT-5.4, GPT-5.5). Guidance comes from an engineered rule library derived from Robert McKee's story theory plus a reference corpus of 31 masterworks (572 annotated scenes); no model weights are changed.
How much does MUSE improve story quality?
Under LLM-judged evaluation, it adds 1.1–6.2 WritingBench points over zero-shot across four base models and raises LongStoryEval by more than ten points on three of the four. Consistency error density stays at 0.245–2.117 per 10,000 words — below reproduced multi-agent baselines on three models, but above plain zero-shot prompting on all four. All scores are LLM-judged; no human evaluation is reported.
Is the code and data available?
Yes. The paper points to https://github.com/RoadtoAGI/MUSE, and appendices document artifact schemas, benchmark query IDs, and a supplementary bundle of generated outputs indexed by query. Rerunning the pipeline still requires paid API access to frontier LLMs and consumes substantially more tokens than zero-shot generation (exact cost is not reported).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction truncated mid-§3.3 and parts of §4; Eqs. 6–8 (context assembly and the review verdict loop) are not visible, so harness internals were assessed from §3.3 prose and Appendix C.
- Figures 1–3 are not analyzable as images; Figure 3's per-domain values were only partially recovered from extracted axis labels.
- Appendices G (dataset statistics/query-id lists), H (per-benchmark baseline rows), and L (serial-fiction extension) are referenced but absent from the extracted text.
- The actual rule library / skill prompts — the claimed core asset — are characterized but not printed in the extracted text, so 'rule quality' claims could not be audited directly.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.