Park, Yang and Kseniia: What Breaks When Facts Become Stories?

Park, Yang and Kseniia: What Breaks When Facts Become Stories?

Kwangsuk Park, JiWoong Yang and Kseniia Vaniushkina argue that source-agreement metrics can misjudge models that turn recorded facts into stories. In reconstruction, departing from the source may be intended: agreement-based evaluation can count warranted changes as errors, while ignoring agreement can let real distortions pass. The authors had three model families rewrite 2,000 news articles as summaries and stories, then assessed changes across the five Ws. They report that time and manner were replaced, while place and causation lost expressions that carry them. Accountability distortion rose up to sevenfold in stories compared with summaries; among cases whose direction could be determined, two in three made the perpetrator seem less responsible. The authors say these distortions can survive claim-level fact checking, leaving a retelling grammatically cleaner than its source. The distinction between a change and a warranted change remains contested: the authors report that three relabeling authors converged on whether distortion occurred, but not whether it was warranted. They release ReTold-2k, a reconstructable corpus, scoring protocol and per-judge baselines kept unaggregated. The abstract does not establish how these results generalize beyond the studied articles and model families. How do the authors determine whether a perspective warrants a change?

Publication: 8670c092-e5d2-460b-85fe-72820335f36b

attribution — uncertain: The source lists the third author as "Vaniushkina Kseniia." The summary reorders this to "Kseniia Vaniushkina," while the title treats "Kseniia" as the surname ("Park, Yang and Kseniia"). The two forms conflict, and the source does not settle the name order. Smallest repair: use the name exactly as listed, or a neutral form such as "Park et al." in the title.

Authors:Kwangsuk Park, JiWoong Yang, Vaniushkina Kseniia

summary — uncertain: The draft never says the models wrote from an assigned perspective, yet it ends by asking whether "a perspective" warrants a change. It also omits that this warrant check set aside a third of detected distortions. It drops a scope condition: the three-author relabeling covered only cases where the judges disagreed. Central terms such as five Ws (who, what, when, where, why), "replaced" versus "lost," and claim-level fact checking are not explained for lay readers. The question sits inside the summary while the question field is empty. Repairs: add the assigned-perspective context, the one-third figure and the judge-disagreement scope, and gloss the five Ws.

asked of each detected distortion whether the assigned perspective warranted it, which set a third of them aside

date — pass: The originalPublished value, 2026-10-29, matches the arXiv v1 submission date. It is not the later feed date.

[Submitted on 29 Oct 2026]

support — pass: Each claim is attributed to the authors and matches the abstract. This includes the sevenfold rise, the two-in-three direction finding, survival of claim-level fact checking, and the ReTold-2k release. The limitations correctly mark what the abstract leaves unknown. The draft attributes no unsupported benefits or forecasts.

Such distortions survive claim-level fact checking and leave the retelling grammatically cleaner than its source.

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the original source.