ACL'23 Distinguished Paper: Discriminative Models May Generalize Better Than Autoregressive Ones

The ACL'23 distinguished paper argues discriminative language models may generalize better than autoregressive ones, with experiments showing ArcaneQA's output distribution skews toward seen data while Pangu remains stable on unseen distributions.

2026-09-24 ~ 2026-09-24 · 3 related posts

• ACL'23 outstanding paper: discriminative LMs may generalize better than autoregressive models — ysu_nlp · 2026-09-24
• Autoregressive vs discriminative models: distribution mismatch drives generalization failure — ysu_nlp · 2026-09-24
• Autoregressive ArcaneQA skews to seen data while discriminative Pangu generalizes — ysu_nlp · 2026-09-24