Critic Training Details and Stability
ziv_ravid · x · 2026-07-11
This part details the implementation of training a critic:
- To make the critic trainable, the authors froze the attention layers and only updated the MoE projections, as gradient explosions primarily originate from attention.
- The critic is updated twice per policy step.
- For multi-turn trajectories, they skip observation tokens in the GAE.
Overall, it explains how to handle numerical instabilities in value models and trajectories during agent/RL training.
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