{
  "id": 7393,
  "url": "https://arxiv.org/abs/2603.10444v2",
  "title": "The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training",
  "summary": "FP4 training promises substantial memory and compute savings for large language models, but remains fragile because blockwise quantization is dictated by extreme activation magnitudes, which inflate dynamic range and compress long-tail signals. We identify a counterintuitive source of this failure: dominant activation outliers are not merely arbitrary sparse events, but are largely induced by a coherent rank-one mean bias, whose direction aligns with the leading anisotropic spectral component. T",
  "authors": "Hengjie Cao, Zhendong Huang, Mengyi Chen, Yifeng Yang, Fang Dong, Anrui Chen et al.",
  "category": "research",
  "topics": "bias-fairness",
  "orgs": null,
  "regions": null,
  "published_at": "2026-03-11T05:59:12.000Z",
  "fetched_at": "2026-07-14T16:33:12.390Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/7393",
  "original_url": "https://arxiv.org/abs/2603.10444v2",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}