Researcher Trains Model From Scratch, Finds RL Fails Without Pretraining
DimitrisPapail · x · 2026-11-08
DimitrisPapail says he trained a model from scratch and realized RL doesn't work without pretraining — pretraining is an implicit prerequisite for RL to succeed, sparking a debate with giffmana about tabular settings.
Related event: Researcher Finds RL Fails Without Pretraining, Sparking Debate(2 posts)→
More from Research
- METR Progress per R&D Dollar Fell 63% Even as Yearly Gains Doubled — sebkrier · 2026-11-08
- GenrobotAI's DAS Dex 2.0 captures 21-DoF hand motion plus 980 tactile points for robot learning — CyberRobooo · 2026-11-08
- fordycelab proposes benchmarking binder-design algorithms with Kd measurements on 24,000-VHH libraries — anshulkundaje · 2026-11-08
- WiC is Not WSD: LLMs err by overthinking fine-grained word senses, AACL 2026 paper finds — Bollegala · 2026-11-08
- Latent Sparse Attention paper: indexing eats 98% of prefill at 10M tokens, LSA speeds one layer 22.7x — burny_tech · 2026-11-08
- Double-exponential fit suggests model time horizons could hit ~80 work-years by end of 2027 — scaling01 · 2026-11-08