STaR-DRO: Stateful Tsallis Reweighting for Group-Robust Structured Prediction
Samah Fodeh, Ganesh Puthiaraju, Elyas Irankhah, Linhai Ma, Srivani Talakokkul, Afshan Khan, Sreeraj Ramachandran, Jordan Alpert, Sarah Schellhorn · 2026
Read STaR-DRO: Stateful Tsallis Reweighting for Group-Robust Structured Prediction by Samah Fodeh, Ganesh Puthiaraju, Elyas Irankhah, Linhai Ma, Srivani Talakokkul, Afshan Khan, Sreeraj Ramachandran, Jordan Alpert, Sarah Schellhorn (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-2604-09737.txt.
Introduction
Fodeh and seven co-authors present a method for ontology-constrained prediction, where outputs stay label-accurate, structurally valid, and evidence-grounded. A design combining XML-style structure, chain-of-thought reasoning, and a self-validation gate targets format drift and ambiguity. Their STaR-DRO method upweights hard groups without suppressing easier ones, unlike DRO. On EPPC Miner, prompt engineering lifts label F1 about 14 points and span F1 about 17 points. STaR-DRO lifts label F1 and cuts cross-entropy about 21 percent against fine-tuning and about 15 percent against standard DRO.
Catalog record
| CREATOR | Samah Fodeh, Ganesh Puthiaraju, Elyas Irankhah, Linhai Ma, Srivani Talakokkul, Afshan Khan, Sreeraj Ramachandran, Jordan Alpert, Sarah Schellhorn |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 157664 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.09737