Survey finds continuous LM progress shifted from guidance methods to few-step generation
LucaAmb · x · 2026-11-09
Writing a survey on continuous language models, @sophiatang observed a clear trend shift: up to 2023, progress was mostly driven by guidance capabilities — using well-defined rewards, classifier gradients, or CFG. Over the past year, advances have instead been driven by the push toward few-step generation, a wave that started in images with methods like flow maps and drifting models.
The survey visualizes this in an interactive timeline (Figure 1) comparing methods for continuous LMs against continuous image/video generation since 2020.
More from Research
- Dev builds singing synth that models vocal tract as branching acoustic tubes, not spliced samples — flngr · 2026-11-09
- DefFlow: Robots Manipulate Ropes, Cloth into Precise 3D Goal Shapes via Point Flows — drsrinathsridha · 2026-11-09
- RLDM 2027 announces Brian Christian, author of The Alignment Problem, as speaker — dabelcs · 2026-11-09
- Nature review: AI has transformed maths — and may outgrow human insight — LucaAmb · 2026-11-09
- No vision, no tactile sensors: humanoid dexterity from proprioception alone at CoRL — Jan_R_Peters · 2026-11-09
- LowRAM: Huawei's KV cache compression gets 4x near-lossless value amortisation — Teresa Delgado · 2026-11-09