The authors compare learned correctness predictors with self-confidence, defined as the probability a model assigns to its chosen answer.
Better language models leave less room for separate routing predictors
Zhang and Hardt connect prediction gains beyond a model’s own confidence to its room for improvement, then test the relationship across 16 language models.
Independent
Guanhua Zhang · Moritz Hardt
Research Digest··2 min read
Guanhua Zhang and Moritz Hardt study when a separate predictor can judge a language model’s answers better than the model’s own confidence.
§