Beyond Logit Adjustment: A Residual Decomposition Framework for Long-Tailed Reranking
Zhanliang Wang, Hongzhuo Chen, Quan Minh Nguyen, Mian Umair Ahsan, Kai Wang, [4pt] University of Pennsylvania, Philadelphia, PA, USA · 2026
Read Beyond Logit Adjustment: A Residual Decomposition Framework for Long-Tailed Reranking by Zhanliang Wang, Hongzhuo Chen, Quan Minh Nguyen, Mian Umair Ahsan, Kai Wang, [4pt] University of Pennsylvania, Philadelphia, PA, USA (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-2604-01506.txt.
Description
Wang and colleagues study long-tailed classification, where few frequent classes dominate many rare ones. Through Bayes-optimal reranking on a top-k shortlist, they show the residual splits into classwise and pairwise parts. When the residual is purely classwise, a fixed offset recovers the ordering, but incompatible constraints across contexts defeat any fixed offset. They develop REPAIR, a lightweight reranker combining a stabilized classwise term with a linear pairwise term from competition features. Nine benchmarks confirm the decomposition explains where pairwise correction helps and where classwise correction alone suffices.
Catalog record
| CREATOR | Zhanliang Wang, Hongzhuo Chen, Quan Minh Nguyen, Mian Umair Ahsan, Kai Wang, [4pt] University of Pennsylvania, Philadelphia, PA, USA |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 59043 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.01506