Confidence signals provide limited help in extractive question answering
Intrinsic Sequence-Likelihood Confidence in Retrieval-Dominated Extractive QA: Two Pre-Specified Negatives, and What They Do and Do Not Attribute
Computation and LanguageInformation RetrievalMachine Learning
Summary
When a computer answers questions by picking pieces from a document, it often already retrieves the best possible answers. The authors studied whether the computer's own confidence in its answers could help decide which answers to trust or which questions to focus on learning more about. They found that this confidence signal doesn't improve performance much, before or after extra training. Instead, simple methods based on comparing some predefined negative examples may offer clearer insights.
What this means in practice
- •For machine learning engineers: Design QA systems that focus on retrieval quality rather than model confidence signals for answer trustworthiness and query selection.
- •For natural language processing tool builders: Develop QA components that use explicit negative examples to better understand model behavior instead of relying on sequence likelihood confidence.
Authors
Gunwoo Lee, Changmin Sung, Sang-Hwan Gwak, Ina Kim, Ji-Young Choi, Kyong-Ha Lee
Abstract
In extractive document question answering whose questions were generated from the passages that contain their answers -- so that retrieval recovers 92-99.8% of what any mode combination could reach, whatever its absolute accuracy -- confidence-driven mechanisms have little to gain. Fine-tuning an open language model on a specialized domain corpus yields a model whose own confidence is a tempting control signal: it could decide which queries warrant further adaptation, and which answers to trust. We evaluate both uses under criteria fixed before the runs were executed, across four 7-9B model families whose adaptation moved closed-book F1 by at most +0.03, and both fail: a distillation trigger on all four families, under its pre-specified three-step transfer budget, and a routing-and-abstention policy in its single-model pilot. Retrieval alone recovers 92-99.8% of best-case combined accuracy under every correctness criterion we test, leaving routers no meaningful gain. The sequence-likelihood signal is insufficient relative to that mode -- area under the receiver operating characteristic curve 0.65-0.81 under the registered criterion -- before adaptation as well as after, unchanged by scalar recalibration and not consistently improved by token-level temperature rescaling. And the finer diagnostics depend on the correctness criterion and on answer length; on the three adapted combinations where we could test it, selector ablations show no statistically detectable downstream benefit from the confidence term on any seed; on Gemma, removing it changes the selector from failing to passing both registered criteria. The usable product is a set of pre-specified negatives with their dependencies made explicit.