https://yiziruifang.com/tags/selective-prediction/Selective PredictionHugo Blox Builder (https://hugoblox.com)en-usSun, 01 Feb 2026 00:00:00 +0000https://yiziruifang.com/media/icon_hu7729264130191091259.pnghttps://yiziruifang.com/tags/selective-prediction/https://yiziruifang.com/publication/fang-2026-learning-tmlr/Sun, 01 Feb 2026 00:00:00 +0000https://yiziruifang.com/publication/fang-2026-learning-tmlr/<h2 id="research-question">Research question</h2> <p>How should a human-AI system act when its learned routing function is itself uncertain about whether the model or human should make a prediction?</p> <h2 id="approach-and-evidence">Approach and evidence</h2> <p>The paper applies conformal prediction to the rejector, replacing a forced binary decision with prediction sets or intervals. It studies two selective workflows: abstaining when the allocation is uncertain, and querying both decision makers to check agreement. Experiments span image and hate-speech classification. The and document the methods and evaluation.</p> <h2 id="scope-and-versions">Scope and versions</h2> <p>The uncertainty estimate concerns assignment of responsibility between a model and an expert. Abstention withholds a prediction; consensus checking requires both predictions, so these workflows change which decisions are made and when the expert is consulted. This is the February 2026 TMLR article; the is an earlier version of the same research.</p>https://yiziruifang.com/publication/fang-2024-learning-workshop/Sun, 01 Dec 2024 00:00:00 +0000https://yiziruifang.com/publication/fang-2024-learning-workshop/<h2 id="research-question">Research question</h2> <p>Can uncertainty about the rejector improve decision making in learning-to-defer systems, where each input is assigned to either a model or a human expert?</p> <h2 id="approach-and-evidence">Approach and evidence</h2> <p>This workshop paper uses conformal prediction to make the rejector return sets rather than a single defer-or-predict decision. It evaluates two responses to ambiguous routing: withholding the prediction and checking agreement between the human and model. The reported experiments cover tasks ranging from object to hate-speech detection. The contains the original method and results.</p> <h2 id="scope-and-versions">Scope and versions</h2> <p>The paper focuses on multiclass learning to defer with one expert. Its selective workflows can abstain or consult both decision makers, so their behavior includes decisions that are withheld as well as those returned. This is the NeurIPS 2024 Workshop on Bayesian Decision-making and Uncertainty version. The subsequent has a separate abstract, review record, and citation.</p>