Sticking with GRPO: The Four Targeted Mods Behind Xiaomi's Stable MoE RL
tokenbender · x · 2026-09-22
tokenbender breaks down Xiaomi's MoE RL recipe: no fancy algorithm swap, just targeted GRPO modifications for stability and scale:
- Prompt-mean loss instead of token-mean to stop long answers from dominating gradients (from DAPO)
- Separate clipping bounds for positive and negative advantages (DAPO)
- Entropy-adaptive clipping (mai-thinking-1)
- Segment-level credit shaping (likely TreeAdv)
Related event: Xiaomi's MoE RL stability paper: router-aware GRPO tweaks explained(2 posts)→
More from Research
- SGLang Megakernel Pushes Muse Spark to 1,109 tok/s, 2.1x Over Baseline — hsu_byron · 2026-11-12
- ArcGrav: a cross-model Agent Control Language to cut token overhead — ArchitecturalGravity · 2026-11-12
- Open-source dual-arm robot solves tangram puzzles with 90 demos, ~60% success — DanGrover · 2026-11-12
- FutureHouse's deslopper trains Kosmos to write indistinguishably from human experts — proteinrosh · 2026-11-12
- Fleuret amplifies debate: AI can't slash wet-lab experiments in unmapped biology — francoisfleuret · 2026-11-12
- Recursive's Eureka Machine ships Flint: a sparse 40B model with 2.5B active for local agentic coding — agihouse_org · 2026-11-12