Improving Language Model Confidence Estimation with Linear Ambiguity Representations

1. 11:52
   arXiv

   Improving Language Model Confidence Estimation with Linear Representations of Ambiguity

   Researchers extract a linear representation of input ambiguity from language model activations and propose the DART fine-tuning method to improve confidence calibration. The authors apply difference-in-means to minimal pairs of ambiguous and unambiguous sentences to obtain an ambiguity direction, and find that steering along this direction weakly but selectively changes the model's judgment of input ambiguity. DART encourages parameter updates to lie within the span of these ambiguity directions and penalizes activation changes outside this subspace. According to the abstract, the method significantly improves calibration, writes the improvement into model parameters, and maintains accuracy on multiple-choice and generative question-answering tasks.

Current attention 27·Peak in the comparable period 30 (Nov 10 23:00)·Change over the comparable past 24 hours -1%