EditPropBench: Measuring Factual Edit Propagation in Scientific Manuscripts
Garvin Kruthof · 2026
Read EditPropBench: Measuring Factual Edit Propagation in Scientific Manuscripts by Garvin Kruthof (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-02083.txt.
Description
Garvin Kruthof introduces EditPropBench, a benchmark for whether language model editors carry a factual edit through a whole scientific manuscript. A local change, such as a data set shrinking from 215 to 80 documents, can make distant claims stale even when they never repeat the number. An audit of recent arXiv papers found such fact-dependent claims in over a third of benchmark papers.
Each benchmark item holds a synthetic manuscript in machine learning style, a target edit, and a fact graph with sentence labels. Labels mark direct targets, required downstream updates, and text that must stay unchanged. A score called Edit-Ripple Adherence measures the share of required updates done right, checked with adversarial probes.
Catalog record
| CREATOR | Garvin Kruthof |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.02083