Study Confirms Mismatch in Async and Group Sampling
ziv_ravid · x · 2026-07-11
A tweet discussed the mismatch between asynchronous processing and group sampling in RLHF, pointing out that this is a practical phenomenon worth deeper investigation.
- Model Performance: Experiments show that GRPO (Group Relative Policy Optimization) collapses quickly during training, whereas SAO remains stable.
- Benchmarks: SAO outperformed GRPO variants on the SWE-Bench Verified and multiple math benchmarks.
- Dynamic Tracking: In an experiment where reward preferences shifted midway, the value model tracked the changes much faster than methods using a running mean as a baseline.
Related event: GLM Team Proposes SAO Algorithm for Asynchronous Agent RL(15 posts)→
More from Research
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27