A Decoding Algorithm for Length-Control Summarization Based on Directed Acyclic Transformers
Chenyang Huang, Hao Zhou, Cameron Jen, Kangjie Zheng, Osmar R. Zaïane, Lili Mou · 2025
Read A Decoding Algorithm for Length-Control Summarization Based on Directed Acyclic Transformers by Chenyang Huang, Hao Zhou, Cameron Jen, Kangjie Zheng, Osmar R. Zaïane, Lili Mou (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-2502-04535.txt.
Description
Chenyang Huang and co-authors present a computing paper from 2025 on summaries with a strict length cap. Many tasks need a summary that fits a set budget, not just a shortish one. Past models treat the cap as a soft wish and often miss it.
Their method builds on a transformer that lays out text pieces as a web rather than a chain. A decoder then picks a path through the web that forms the summary. A new search scores whole paths at once, weighing all ways the pieces could join. Beam search plus a reranker polish the final pick. On two news data sets, the method set the best scores for cap-respecting summaries. The authors say hard caps need search over paths, not token-by-token writing.
Catalog record
| CREATOR | Chenyang Huang, Hao Zhou, Cameron Jen, Kangjie Zheng, Osmar R. Zaïane, Lili Mou |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-04535