Ablation thread: full-trajectory reward sampling beats early- or late-step-only ReFL training
gowthami_s · x · 2026-10-17
A thread analyzing a video generation paper using ReFL (reward feedback learning) points out a missing comparison: the same training setup with a different full-video reward model, which would isolate how much of the gain comes from the generator backbone itself. The paper's ablations show that with the same reward model, training only early denoising steps scores better on motion while late-only training does better on anatomy, and sampling steps across the full trajectory scores best on both — suggesting distinct feedback signals at different diffusion stages.
Related event: PRFL: Latent Process Reward Feedback Speeds Up Video Generation Alignment(5 posts)→
More from Research
- Terry Tao's Four Types of Mathematization: Heavy AI Use Needs Strong Human Understanding — ravithejads · 2026-10-17
- Princeton's LabOS, Featured in Nature: An AI Co-Scientist for Hypothesis and Experiment Design — MengdiWang10 · 2026-10-17
- 'Tip-of-the-tongue' queries: a blind spot for modern agentic retrieval — bclavie · 2026-10-17
- Tip-of-the-tongue queries emerge as a key test of agentic retrieval systems — ctjlewis · 2026-10-17
- micro1 launches LegalReasonBench to test models on attorney-authored legal matters — douwekiela · 2026-10-17
- LEASH: weak-to-strong escalation matches Claude Opus 5 at 19.2% lower cost on OSWorld — xwang_lk · 2026-10-17