SAO Outperforms GRPO Variants
ziv_ravid · x · 2026-07-11
This thread discusses a new RL training approach and shares experimental results:
- Is SAO slower/more stable to converge than GRPO? The original text clarifies that the authors found GRPO degrades faster, while SAO does not, and that SAO outperforms GRPO variants on SWE-Bench Verified and math benchmarks.
- They also conducted an experiment: when reward preferences changed midway through training, the value model tracked the new preferences faster than a "running-mean baseline."
This post is essentially a sharing of research observations regarding agent/RL training stability and the role of value models.
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