The Alignment Bottleneck in Decomposition-Based Claim Verification
Mahmud Elahi Akhter, Federico Ruggeri, Iman Munire Bilal, Rob Procter, Maria Liakata · 2026
Read The Alignment Bottleneck in Decomposition-Based Claim Verification by Mahmud Elahi Akhter, Federico Ruggeri, Iman Munire Bilal, Rob Procter, Maria Liakata (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-2602-10380.txt.
Description
Breaking complex claims into sub-claims is often proposed as a verification fix, yet results have been inconsistent. A 2026 paper led by Mahmud Elahi Akhter, with four co-authors, blames two overlooked bottlenecks: poor evidence alignment and sub-claim error patterns. They build a dataset of real-world complex claims with temporally bounded evidence. Annotated spans link each sub-claim to its evidence. They test two setups: sub-claim aligned evidence, giving each sub-claim its own evidence, and repeated claim-level evidence, reusing the whole claim's evidence throughout.
Decomposition helps substantially only with fine-grained evidence tied to each sub-claim. Standard setups reusing claim-level evidence fail to improve and often degrade performance. The pattern holds across PHEMEPlus, MMM-Fact, and COVID-Fact. When sub-claim labels are noisy, the authors show, the error's nature shapes downstream robustness. Pipelines must therefore mind both alignment and label quality.
Catalog record
| CREATOR | Mahmud Elahi Akhter, Federico Ruggeri, Iman Munire Bilal, Rob Procter, Maria Liakata |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-10380