CMU researchers propose unified training: one algorithm for the full foundation model pipeline
DanielKhashabi · x · 2026-10-30
Researchers from CMU, Johns Hopkins, and Aurora Innovation (Yuda Song, Alvin Zhang, Fahim Tajwar et al., advised by Jeff Schneider, Ruslan Salakhutdinov, Andrea Zanette, Aarti Singh) propose unified training — a single algorithm spanning the full foundation model pipeline — questioning whether the pretraining/midtraining/post-training split is fundamental.
Method highlights:
- The data mixture combines static data (pretraining, midtraining, SFT) with reward-weighted RL generations; the trainer uses plain maximum likelihood, with generate-and-evaluate closing the loop.
- Claimed benefits: simpler infrastructure and a cleaner framework for training science.
- The accompanying figure shows unified vs. staged accuracy/coverage curves under matched cumulative compute.
First post of a thread; full details in the follow-ups.
Related event: CMU and Collaborators Propose Unified Training with a Single Algorithm(4 posts)→
More from Research
- Healthcare RL tasks are mostly billing code — where are the House-style diagnostic sims? — nirmal_dist · 2026-10-31
- Steering an "Einstein neuron" in Qwen 4B produces dramatically different answers — thursdai_pod · 2026-10-31
- Protein binder scoring head-to-head: ipDAE consistently strong, OpenDDE+ipTM surprisingly good at 80% PPV — AllThingsApx · 2026-10-31
- Turing TTS benchmark: best voice AI fools listeners only ~10% of the time — omarsar0 · 2026-10-31
- Ablation study: fine-grained MoE is the best LLM architecture change since GPT-3 — stochasticchasm · 2026-10-31
- Yann LeCun to Keynote MAIN2026 on World Models for Reasoning and Planning — karimjerbineuro · 2026-10-31