Uncertainty-aware Language Modeling for Selective Question Answering

We present an automatic large language model (LLM) conversion approach that produces uncertainty-aware LLMs capable of estimating uncertainty with every prediction. Our approach is model- and data-agnostic, is computationally-efficient, and does not rely on external models or systems. We evaluate converted models on the selective question answering setting -- to answer as many questions as possible while maintaining a given accuracy, forgoing providing predictions when necessary. As part of our results, we test BERT and Llama 2 model variants on the SQuAD extractive QA task and the TruthfulQA generative QA task. We show that using the uncertainty estimates provided by our approach to selectively answer questions leads to significantly higher accuracy over directly using model probabilities.

Capsa: A Unified Framework for Quantifying Risk in Deep Neural Networks

The modern pervasiveness of large-scale deep neural networks (NNs) is driven by their extraordinary performance on complex problems but is also plagued by their sudden, unexpected, and often catastrophic failures, particularly on challenging scenarios. Existing algorithms that provide risk-awareness to NNs are complex and ad-hoc. Specifically, these methods require significant engineering changes, are often developed only for particular settings, and are not easily composable. Here we present capsa, a framework for extending models with risk-awareness. Capsa provides a methodology for quantifying multiple forms of risk and composing different algorithms together to quantify different risk metrics in parallel. We validate capsa by implementing state-of-the-art uncertainty estimation algorithms within the capsa framework and benchmarking them on complex perception datasets. We demonstrate capsa's ability to easily compose aleatoric uncertainty, epistemic uncertainty, and bias estimation together in a single procedure, and show how this approach provides a comprehensive awareness of NN risk.