RL training stability tricks: prompt-mean loss, asymmetric clipping and entropy-adaptive GRPO
tokenbender · x · 2026-09-22
The author breaks down key RL training modifications: sticking with GRPO rather than fancier algorithms but tuning it for stability and scale — prompt-mean loss over token-mean loss to stop long answers from getting excess gradient weight (DAPO); separate clipping bounds for positive/negative advantages (DAPO); entropy-adaptive clipping (mai-thinking-1); and segment-level credit shaping (TreeAdv).
They also note that 13% policy-drift staleness with large throughput implies the chosen environments run or improve directly on base weights without any curriculum — otherwise training would have suffered penalties.
More from Models
- Datalab claims parsing accuracy second only to Gemini at 1/37 the cost and 6x speed — VikParuchuri · 2026-11-12
- Datalab fast-mode OCR beats Opus 5.5 and GPT-6.1 Sol at 1/4 to 1/68 the price — VikParuchuri · 2026-11-12
- Claude's world knowledge impresses: model guesses user identity from one prompt — adonis_singh · 2026-11-12
- Gemini 4 Carbon opens to all at $2/$10, price doubling to $4/$20 on Dec 1 — socialwithaayan · 2026-11-12
- Ten wild Gemini 4 Carbon demos in 24 hours: CUDA meadows, V8 engines, mech stomps — socialwithaayan · 2026-11-12
- How to cut Fable 5.2 token costs by 90%: Arc routing plus a lean CLAUDE.md — EXM7777 · 2026-11-12