Sequence-level supervised learning for AR LMs is fundamentally flawed, Google team shows

mblondel_ml · x · 2026-11-17

Mathieu Blondel et al.'s arXiv paper thread (full summary in the first post): teacher forcing causes exposure bias, but the two natural sequence-level losses (greedy decoding and AR sampling based) are provably unbounded below; a stop-gradient variant yields a valid but high-variance moment matching direction. Conclusions: teacher forcing stays, OPD is better but needs a teacher, sequence-level training remains RL's job.

Related event: DeepMind paper proves sequence-level supervised losses for autoregressive models can be fundamentally flawed(5 posts)→

Original post →

More from Research

Research channel →