New QUADS technique stabilizes NVFP4 RL for MoE LLMs

Researchers have developed a new technique called QUADS to stabilize reinforcement learning (RL) for Mixture-of-Experts (MoE) Large Language Models using the NVFP4 low-precision format. They identified activation error, not weight error, as the primary cause of instability in FP4 RL for MoE models. QUADS addresses this by implementing asymmetric quantization-aware training on the trainer side and residual activation compensation on the rollout side, achieving BF16-level accuracy and improving throughput. AI

IMPACT This research could lead to more efficient training of large language models by enabling the use of lower-precision formats without sacrificing accuracy.

RANK_REASON This is a research paper detailing a novel technique for stabilizing reinforcement learning in MoE LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• E2M1
• FP8
• Hugging Face
• Large Language Models
• Mixture-of-Experts
• NVFP4
• QUADS
• reinforcement learning
• W4A4

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →