Researchers break down RFD4-Proteina's CFM trick: recurrent hidden-state refinement, fewer ODE steps
AllThingsApx · x · 2026-10-22
Anirban Ray highlights the interesting CFM idea behind RFD4-Proteina (Baker Lab × NVIDIA GenAIR): keep xt fixed and recurrently refine hidden states, with ODE sampling particularly well suited. Compared to SDE sampling, ODE needs fewer steps, leaving more compute for reasoning within each step — more reasoning per step, fewer steps overall.
More from Research
- Claimed unpatchable PyTorch autograd zero-day runs arbitrary code via SGD — burny_tech · 2026-10-22
- Research Progress Graph: modeling authentic research trajectories, 35 agents found incomplete — ZhiyuChen4 · 2026-10-22
- EMNLP 2026 paper uses prediction-powered inference to quantify how much human annotation LLM judges save — shi_weiyan · 2026-10-22
- Ask-to-Distill: synthesizing questions for answer-free on-policy distillation — UIUC-CS · 2026-10-22
- 2-bit draft in VRAM + exact residual in RAM: speculative decoding for models that don't fit your GPU — Happy_Guitar3521 · 2026-10-22
- RFD4 team swaps self-conditioning for recurrent updates, echoing looped flows — AllThingsApx · 2026-10-22