Source-linked AI summary
PAUSE: Editable Strategy Artifacts for Long-Form Cultural Story Adaptation
Taaha Kazi, Vasu Sharma, Mohammad Saifullah, Abdur Rahman
TL;DR
PAUSE addresses the limited visibility of cultural decisions in long-form AI adaptation by exposing the model’s intermediate strategy for human editing. It implements pause/edit/resume propagation through localization and chapter-writing stages, and finds consistent edit adherence in a smoke-scale study. The evidence supports inspectability and contestability, but remains limited in scope and does not establish cultural adequacy or literary improvement.
Problem
Cultural decisions in long-form AI adaptation often remain hidden, even though names, settings, institutions, hierarchies, and genre register must remain stable across chapters.
Method
PAUSE emits a structured adaptation strategy, exposes it as an editable JSON artifact, and resumes downstream character, entity, and chapter-localization stages from the edited copy.
Results
Across 9 edited-vs-control chapter comparisons, judges selected edited-strategy outputs in all 9, with target markers in 8/9 edited outputs versus 0/9 controls.
Takeaways & Limitations
PAUSE provides a practical control surface for making AI-mediated cultural adaptation more inspectable, controllable, and contestable before decisions reach prose.
Takeaways & Limitations
The evaluation covers two Chinese-source novels and Chinese-to-American-English localization, so broader stories and language-culture pairs are needed to test generalization.
Abstract
from arXiv · showhide
Generative AI systems increasingly mediate cultural adaptation, but their cultural decisions are often hidden inside prompts, transient model plans, or final prose. We study PAUSE (Pause-And-Update Strategy Editing), an intervention that exposes an editable adaptation strategy as a human control surface for cultural decisions in long-form story adaptation. The strategy is a structured artifact that can be inspected, edited, and then projected through downstream character, entity, and chapter-localization stages. In two Chinese-source serialized novels, we test whether human edits to this strategy propagate into chapter-level prose. Across 9 edited-vs-control chapter comparisons, judges select the edited-strategy output in all 9; a marker audit shows target markers in 8/9 edited outputs and 0/9 controls, with forbidden markers absent from edited outputs and present in all controls. We frame these results as a smoke-scale edit-adherence study, not a claim that the outputs are culturally authoritative or literary-quality improvements. PAUSE offers one practical way to make AI-mediated cultural adaptation more inspectable and contestable before decisions propagate through long-form generation.
1. Introduction
PAUSE exposes an editable intermediate adaptation strategy so cultural decisions can be inspected and revised before reaching long-form chapter prose. A smoke-scale study tests whether those edits propagate, while explicitly limiting claims about cultural adequacy and literary quality.
- Generative AI cultural decisions often remain hidden in prompts, transient plans, or final prose, complicating long-form adaptation across stable names, settings, institutions, hierarchies, and genre register.
- PAUSE exposes the model-generated adaptation strategy as an editable JSON artifact covering target setting, naming conventions, institutional analogues, register choices, and entity-localization principles.
- Across 9 chapter-level comparisons, judges selected edited-strategy outputs in all 9; target markers appeared in 8/9 edited outputs and 0/9 controls.Forbidden markers appeared in 0/9 edited outputs and 9/9 controls.
- The paper frames the evidence as smoke-scale edit adherence showing propagation into prose, not cultural authority, literary-quality improvement, or full-novel coherence.
- PAUSE contributes a pause/edit/resume control surface with direct, character-mapping, and entity-mapping paths from strategy edits into prose.
2. Related Work
Related work situates PAUSE at the intersection of long-form literary translation, intermediate writing artifacts, and narrow evaluation of open-ended generation.
- Long-form literary translation and cultural adaptation: Long-form literary translation research motivates consistency beyond individual sentences or scenes, including discourse-level Chinese–English and chapter-to-chapter context.
- Human control through intermediate writing artifacts: Human-AI writing systems expose text-level controls, while long-form generation systems use plans, outlines, memories, and revision loops as intermediate guidance objects.
- Evaluation of open-ended generation: Because creative and cultural adaptation lacks one reference answer, PAUSE evaluates edit adherence with pairwise judgments and a deterministic marker audit rather than global quality.
3. Pipeline Substrate
The pipeline separates metadata localization from chapter writing and makes the strategy-edit claim boundary explicit: edits influence prose through three strategy-derived paths before a strategy-blind polish pass.
- Metadata layer: The metadata layer extracts chapter entities and relationships, condenses them into a structured strategy, localizes characters and entities, and validates the resulting records.
- Metadata layer: The strategy JSON contains declarative rules for names, families, places, institutions, social address, genre terms, and cultural references, with a source-conditional schema.
- Localized records and mention projection: Character localization creates source-to-target names and mention maps, while entity localization adapts related graph segments together before records are projected into text processing.
- Text layer: Chapter writing first adapts and replaces mentions, then rewrites the draft for fluency and style; the polish pass receives no strategy JSON directly.
- Claim boundary: Strategy edits can reach prose through direct strategy text, character mappings, and entity mappings, but they do not control every sentence.
4. PAUSE: Editable Strategy Steering
PAUSE places human intervention between prompt-time steering and final-prose postediting by pausing after strategy generation, editing the emitted artifact, and resuming downstream localization.
- PAUSE exposes the generated strategy, lets an editor revise it, and resumes without fine-tuning the model or requiring whole-output rewriting.
- Pause/edit/resume hook: The prototype writes strategy JSON after extraction and the strategy call, halts for editing, then reloads the edited file to continue localization and chapter writing.
- What is editable: An edit is any change to the emitted JSON document between pause and resume, such as reframing Love and Strings from a New York piano family to a Nashville guitar family.
- What is editable: The same edit can also alter surname conventions, institutions, and honorifics, but whether it reaches final prose requires empirical evaluation.
5. Evaluation
PAUSE is evaluated through controlled edited-versus-control chapter comparisons, pairwise judgments, and a deterministic marker audit. Across nine chapter comparisons, the results support strategy-edit propagation into prose while remaining a smoke-scale edit-adherence study.
- Each comparison paired a control resumed from the unedited strategy with an edited localization resumed after a human strategy edit.The source, captured baseline strategy, downstream configuration, and edit description were held constant except for the intended strategy edit.
- Across all 9 judged chapters, judges selected the edited-strategy output in all 9 comparisons.The runs covered two Chinese-source serialized novels and resumed edited and control localizations from the same captured baseline strategy.
- The study is framed as smoke-scale edit adherence rather than evidence of cultural adequacy, literary quality, or full-novel coherence.The evaluation uses live model aliases without a fixed seed, so chapter rewriting is nondeterministic in absolute output, although each comparison shares a captured baseline and downstream configuration.
- Forbidden baseline markers appeared in 0/9 edited outputs and 9/9 controls.The separation was complete: every control retained at least one forbidden marker, whereas no edited output did.
- The secondary rubric audit produced 18/18 edited-side selections, with source faithfulness tied and target-culture naturalness and prose quality similar.All nine chapters had edited-side wins under both shuffled presentation orders; the audit was treated as secondary evidence rather than cultural correctness.
6. Discussion and Limitations
PAUSE makes the intermediate cultural plan visible, editable, and testable before it propagates into long-form prose, while the evaluation remains a smoke-scale study limited to two novels and one language-culture direction.
- PAUSE exposes the model’s intermediate cultural plan as a visible, editable, and testable control surface before prose generation.
- The study covers two source novels and Chinese-to-American-English localization, leaving generalization to more stories and language-culture pairs for future work.
- The pause-and-resume hook lets human edits to the strategy propagate into chapter prose through the long-form localization workflow.
- 9 chapter-level comparisons across two Chinese-source web novels support the reported edit-adherence result through pairwise judgments, marker audits, and shuffled re-judging.
A. A worked example of an edit
The worked example shows how an editor reframes a family’s cultural setting and milieu-specific titles within an existing strategy, then projects those changes through downstream localization into chapter writing.
- The example reframes the Ashford family from New York classical piano to Nashville country guitar by editing the adaptation strategy.
- The naming edit changes the consistent family scheme from Ashford to Branson while preserving the foreign-character rule’s narrative function.
- The honorifics-and-titles edit rewrites entries carrying classical-music connotations and leaves the remaining title-mapping rubric unchanged.
- The revised titles replace classical-music terms such as “The Maestro” with culturally different entries including “The Legend” and “Mr. Branson.”
- Both edits remain prose values under existing strategy keys, and their decisions are projected through character, entity, mention, and chapter rewriting stages.