Reproduce “Inducing Artificial Uncertainty in Language Models” on OpenResearch

About this paper

Sophia Hager, Simon Zeng, Nicholas Andrews

Researchers from Johns Hopkins University and Microsoft introduced methods to induce artificial uncertainty in large language models using readily available "easy" data, enabling the training of uncertainty quantification (UQ) probes that effectively generalize to genuinely challenging, unseen data. The techniques, particularly dropout, led to improved confidence calibration on difficult datasets and validated an unsupervised hyperparameter selection approach.

Key insights

• Artificial uncertainty methods (dropout and unlearning) enable UQ probes trained on 'easy' data to effectively generalize and quantify 'real' uncertainty in an original LLM facing 'hard' data.
• Improvements in UQ performance primarily stem from significant reductions in Expected Calibration Error (ECE), enhancing confidence alignment rather than discrimination ability (AUROC).
• Dropout is a robust and computationally efficient method, particularly when applied to attention layers, and is suggested as the more practical choice for inducing artificial uncertainty.

Summary by alphaXiv.