Inducing Artificial Uncertainty in Language Models

TL;DR

This paper explores methods to artificially induce uncertainty in language models to improve their calibration, especially when real challenging data is scarce or unavailable.

Contribution

It introduces the problem of artificial uncertainty induction and demonstrates that training probes on artificially uncertain data enhances real uncertainty recognition.

Findings

Probes trained on artificial uncertainty outperform those trained without it.

Artificial uncertainty training improves calibration on hard data.

Minimal performance loss on easy data when using artificial uncertainty methods.

Abstract

In safety-critical applications, language models should be able to characterize their uncertainty with meaningful probabilities. Many uncertainty quantification approaches require supervised data; however, finding suitable unseen challenging data is increasingly difficult for large language models trained on vast amounts of scraped data. If the model is consistently (and correctly) confident in its predictions, the uncertainty quantification method may consistently overestimate confidence on new and unfamiliar data. Finding data which exhibits enough uncertainty to train supervised uncertainty quantification methods for high-performance models may therefore be challenging, and will increase in difficulty as LLMs saturate datasets. To address this issue, we first introduce the problem of inducing artificial uncertainty in language models, then investigate methods of inducing artificial uncertainty on trivially easy data in the absence of challenging data at training time. We use probes trained to recognize artificial uncertainty on the original model, and find that these probes trained on artificial uncertainty outperform probes trained without artificial uncertainty in recognizing real uncertainty, achieving notably higher calibration on hard data with minimal loss of performance on easy data.