Reproduce “Selective Question Answering under Domain Shift” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Amita Kamath, Robin Jia, Percy Liang

A study from Stanford University proposes and evaluates selective question answering under domain shift, where models must abstain from answering questions when likely to be incorrect, especially on out-of-domain data. Their calibrator, trained with auxiliary out-of-domain data, achieved 56.1% coverage at 80% accuracy, surpassing MaxProb baselines, and effectively mitigates model overconfidence on novel out-of-domain inputs.

Key insights

  • Base QA models tend to be overconfident on out-of-domain inputs, causing standard confidence metrics to make insufficient abstention decisions in mixed data environments.
  • Leveraging a small amount of *known* out-of-domain data to train a separate calibrator is more effective for improving selective prediction under domain shift than directly augmenting the base QA model's training data.
  • Even when *known* OOD data differs from the *unknown* OOD test data, it still provides valuable signal for a calibrator to generalize better to novel distribution shifts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.