Autoregressive vs discriminative models: distribution mismatch drives generalization failure

ysu_nlp · x · 2026-09-24

A comparison study shows autoregressive models (ArcaneQA) skew heavily toward seen data, creating distributional mismatch and generalization failure from seen to unseen cases, while discriminative models (Pangu) produce much better-aligned output probability distributions over seen and unseen data. Paper linked in the post.

Related event: ACL'23 Distinguished Paper: Discriminative Models May Generalize Better Than Autoregressive Ones(3 posts)→

Original post →

More from Research

Research channel →