RL alone stalls at chance: Carbon-X shows midtraining unlocks gains RL can't
LoubnaBenAllal1 · x · 2026-10-22
The team shares a key methodological finding: applying RL directly to Qwen3-14B never rose above the 50% random-chance baseline on FrontierVEP throughout training, while the same RL approach starting from the midtrained checkpoint moved beyond chance and kept improving — midtraining unlocked gains that RL alone did not deliver.
Related event: Carbon-X Training: Midtraining Beats Pure RL(2 posts)→
More from Research
- WisPaper AI research agent turns one question into an 18-paper review with verified DOIs in 30 minutes — socialwithaayan · 2026-10-22
- Columbia paper: reward misspecification breaks CoT monitoring; new training lifts detection 18%→85% — burny_tech · 2026-10-22
- AGZO paper: activation-guided zeroth-order optimization beats MeZO, fine-tunes Qwen3-4B where first-order OOMs — burny_tech · 2026-10-22
- Zeroth-order optimization papers need retesting on Dust; pretraining is the holy grail — burny_tech · 2026-10-22
- Edinburgh team's 11 papers: self-play lifts Qwen2.5-7B open-ended score from 24.4 to 34.8 — PMinervini · 2026-10-22
- MIT's Juan Antolin-Diaz to present Attention-Based Vector Autoregressions at AMLEDS — TaniaBabina · 2026-10-22