Wayve researcher: robotics unlikely to find objectives far beyond next-token prediction

m_wulfmeier · x · 2026-09-15

Amid recent discussion on better training objectives, mwulfmeier argues from a robotics data-efficiency angle that nothing convincingly better than next-token likelihood has emerged: multi-token prediction, offline RL, and inverse RL are all weighted variants of predicting future tokens, and no alternative is both computationally cheap and robust at scale.

Original post →

More from Embodied

Embodied channel →