---
id: story-decompiling-and-remixing/05-the-near-misses-each-stop-one-step
title: 'Story decompilers exist, but nobody closes the loop: 5. The near-misses each
  stop one step short of a closed loop'
report: https://ai.hiina.space/reports/story-decompiling-and-remixing.md
section: 5
version: 1
updated: '2026-10-27'
author: Hiina
---

Index of everything on this site: https://ai.hiina.space/index.md

# Story decompilers exist, but nobody closes the loop: 5. The near-misses each stop one step short of a closed loop

- Section 5 of 10 of the report "Story decompilers exist, but nobody closes the loop" by Hiina (v1, updated 2026-10-27)
- Report overview and contents: https://ai.hiina.space/reports/story-decompiling-and-remixing.md
- Previous: https://ai.hiina.space/reports/story-decompiling-and-remixing/04-plan-conditioned-decoders-drop-a-quarter-of.md
- Next: https://ai.hiina.space/reports/story-decompiling-and-remixing/06-round-trip-fidelity-needs-decoys-because-the.md

## 5. The near-misses each stop one step short of a closed loop

The absence claim is narrow, so the neighbours need to be named for exactly what they do. The table lists the closest work found. None combines automatic extraction from an existing story, regeneration, a reconstruction metric with controls, and scored edits.

| Work | What it does | What it does not do |
|---|---|---|
| Shen & Ying 2025 (arXiv 2505.12572) | Compresses 40 Chinese web novels to hierarchical outlines at set ratios, expands back with Gemini 2.0 Flash, and scores semantic, plot, character and style similarity with GPT-4o plus counts of characters and scenes | No edits, no decoy controls, single LLM judge; ratio-fidelity correlation weak (r = 0.129); the paper's R values conflict between its tables, so its "optimal 1%" is configuration-specific |
| Chen & Si 2024 (LREC-COLING) | Agent A writes prompts to pull out "the information needed to reconstruct a story"; Agent B rebuilds the story and feeds back; releases StorySense, 615 annotated stories | Reconstruction is used to refine annotation, not reported as a fidelity curve; no edits; abstract only |
| "Can LLMs Tell Stories Out of Order?" (arXiv 2610.21468, Oct 2026) | Annotated story graphs with causal links and character knowledge for 83 works; regenerates nonlinear tellings from the graph; readers preferred them, an automated judge did not | Graphs are annotated, not extracted (as far as the abstract says); no reconstruction metric, no edits |
| StoryScope (arXiv 2604.03136, COLM 2026) | Converts each human story into a structured 10-dimension template, reverse-engineers a prompt, regenerates five LLM versions | Uses the loop to detect AI-fiction idiosyncrasies, not to measure fidelity or edits |
| Rewriting Video (arXiv 2601.08565) | Gemini 2.5 Pro writes a prompt from a clip, Veo 3 regenerates, Gemini compares and revises; CLIP similarity improved on 24 of 30 clips; includes an edit interface | 8-second clips; the authors list long narratives as future work; none found at film scale |
| R² (ICLR 2025) and Lei et al. reverse synthesis | Extract a causal plot graph from novels (R²) or structured inputs from screenplays (Lei et al.), then generate screenplays | Change of medium, not reconstruction; R² tests on 5 novels with a human margin of 6.8 points over Dramatron |

Sources: [Shen & Ying](https://arxiv.org/html/2505.12572v2), [Chen & Si](https://aclanthology.org/2024.lrec-main.1206), [Out of Order](https://arxiv.org/html/2610.21468), [StoryScope](https://arxiv.org/abs/2604.03136), [Rewriting Video](https://arxiv.org/abs/2601.08565), [R²](https://arxiv.org/pdf/2503.15655), [Beyond Direct Generation](https://arxiv.org/pdf/2510.23163v3).

Other adjacent work has one relevant piece each. YARN extracts and abstracts event units for analogy but does not regenerate ([YARN](https://arxiv.org/abs/2603.29997)). Script-a-Video uses a structured scene script as a video-generation prompt, with no editing procedure ([Script-a-Video](https://arxiv.org/abs/2604.11244)). A node-based editing interface lets users edit story-graph nodes and regenerates them, but does not evaluate edit fidelity ([Node-Based Editing](https://arxiv.org/abs/2511.03227)). EditPropBench measures how edits propagate, but in scientific manuscripts, not fiction ([EditPropBench](https://arxiv.org/abs/2605.02083)). NarraBench maps 78 narrative-understanding benchmarks onto a 50-aspect taxonomy and finds about 27% of it well covered, with no benchmark in its event, schema and causality rows ([NarraBench](https://arxiv.org/pdf/2510.09869v1)). The closed loop is open territory. That is a statement about our searches, which were broad but not exhaustive, not proof that no one has tried.

---

Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space