Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection

In 30 Sekunden

Was passiert ist

Researchers trained language models and separate answer selectors to decide when to abstain from answering factual questions rather than guess incorrectly.

Warum es zählt

Matters for engineers building QA systems where wrong answers harm more than admitting uncertainty, especially on factual retrieval tasks.

Achtung

Results vary significantly across datasets. TriviaQA shows improvement but NQ-Open still favors always abstaining, suggesting the approach lacks generalization.