1 Pocket FM
ICML 2026 CultureXAI Workshop
PAUSE exposes the model's adaptation strategy as an editable artifact. A human edit reaches the final prose only through three propagation paths; the final polish pass is strategy-blind.
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 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.
A whole book needs consistent names, places, and register, yet AI keeps those cultural decisions inside the model. Editors can only nudge the prompt or fix the prose, never the decisions themselves.
Pairwise wins
Judges pick the edited output in all 9 judged chapters.
Marker audit
Edited keeps the new markers and drops the old; controls do the opposite.
Secondary rubric audit: edit-adherence moves sharply while source faithfulness stays tied, so edits change only what they should.
Each comparison shares one source chapter and the same baseline strategy; the only intended difference is the human edit. A blind pairwise judge picks the output that better follows the edit.
Editing a few strategy fields reframes a New York classical-piano family as a Nashville country-guitar family; those choices are carried through character records, entity records, and the rewrite.
@inproceedings{kazi2026pause,
title = {PAUSE: Editable Strategy Artifacts for Long-Form Cultural Story Adaptation},
author = {Kazi, Taaha and Sharma, Vasu and Saifullah, Mohammad and Rahman, Abdur},
booktitle = {ICML 2026 CultureXAI Workshop},
year = {2026}
}