Avoiding Knowledge Edit Skipping in Multi-hop Question Answering with Guided Decomposition
Yi Liu, Xiangrong Zhu, Xiangyu Liu, Wei Wei, Wei Hu · 2025
Read Avoiding Knowledge Edit Skipping in Multi-hop Question Answering with Guided Decomposition by Yi Liu, Xiangrong Zhu, Xiangyu Liu, Wei Wei, Wei Hu (2025) 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-2509-07555.txt.
Description
Yi Liu and 4 co-authors tackle a failure mode of knowledge editing in language models in a 2025 paper. Knowledge editing updates a model's facts without retraining, often by retrieval-augmented generation, where edited facts are stored and fetched at question time. Such methods handle simple edits well but stumble on multi-hop questions, which chain several facts, because the system skips the relevant edited fact, a problem named edit skipping.
Their method, IRAKE, adds guided decomposition: the question is broken into steps guided by single edited facts and by whole edited cases. That guidance aligns the model's problem-solving granularity with the stored facts, so the right edit is actually used. Experiments show IRAKE reduces edit-skipping failures and beats leading methods on multi-hop question answering after edits. Reliable updating matters as model knowledge ages quickly in a fast-changing world.
Catalog record
| CREATOR | Yi Liu, Xiangrong Zhu, Xiangyu Liu, Wei Wei, Wei Hu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-07555