Rescuing double robustness: safe estimation under complete misspecification

Researchers have proposed new solutions to address issues with doubly robust estimators and confidence estimation in large language models, according to two recent papers on arxiv.org^[1][2].

Double robustness is a key feature of semiparametric and missing data methodology, but doubly robust estimators can perform poorly when all nuisance functions are misspecified, as noted in a paper submitted to arxiv.org in September 2025^[1]. The authors propose a solution based on adaptive correction clipping (DR+ACC), which inherits the favorable properties of doubly robust estimators under correct nuisance specification and prevents instability caused by compounding product of errors. Meanwhile, another paper submitted to arxiv.org in January 2026^[2] highlights the importance of confidence estimation in large language models, arguing that existing evaluations mainly focus on alignment between confidence and correctness, ignoring language variability. The authors propose a framework consisting of three properties: robustness, stability, and sensitivity, to address this issue. They also note that common confidence estimation methods often fail to distinguish semantically different answers, potentially due to ineffective use of generation-side information.