RL training in practice: "batch size 4" really means ~80 trajectories per step
cephaloform · x · 2026-09-14
The author is running a trajectory-sampling RL-style training setup and notes that "batch size 4" actually translates to 80 individual trajectories per step. A batch size 4, 10-step test on a reference task shows promising (possibly noisy) results, and they plan to scale up to batch size 16 (320 trajectories).
More from Research
- Should arXiv reject an AI-discovered cancer breakthrough that passes clinical trials? — IanArawjo · 2026-09-14
- cESA: judging multiple translations at once cuts human eval time and noise — zouharvi · 2026-09-14
- Framing human eval as a bandit problem focuses annotation budget on top models — zouharvi · 2026-09-14
- Pearmut and two eval methods replace Appraise-era tooling at WMT, headed to EMNLP — zouharvi · 2026-09-14
- Looped Flows: diffusion-style denoising plus recurrent reasoning hits 58.8% on ARC-AGI-1 — LucaAmb · 2026-09-14
- rankECE: a ranking-based measure that beats binned ECE for calibration error — _onionesque · 2026-09-14