CMU and Collaborators Propose Unified Training with a Single Algorithm

CMU, Johns Hopkins, and Aurora Innovation propose "unified training," a single likelihood-based algorithm covering the full pipeline from pretraining and SFT to RL post-training, challenging the conventional multi-stage paradigm by simply switching data.

2026-10-30 ~ 2026-10-31 · 4 related posts

• CMU paper: one algorithm for pretraining and post-training — unified training beats staged pipelines — heghbalz · 2026-10-30
• Unified training: one maximum-likelihood algorithm for pretraining through RL — adityakusupati · 2026-10-31

2 near-duplicate retellings: DanielKhashabi · CarnegieMellonU