{"slug": "softmax-reparameterization-for-output-head-quantization", "title": "Softmax Reparameterization for Output-Head Quantization", "summary": "A post-training method called softmax reparameterization selects a functionally equivalent output head before quantization to cut the inference cost of large-vocabulary output heads in small language models, according to the researchers behind the proposal. The technique subtracts a scalar multiple of the vocabulary-row mean from every output-head row, leaving the softmax distribution unchanged while making the head easier to quantize.", "body_md": "Large vocabularies make output heads a substantial inference cost in small language models. We propose softmax reparameterization, a post-training method that selects a functionally equivalent output head before quantization. The method subtracts a scalar multiple of the vocabulary-row mean from eve", "url": "https://wpnews.pro/news/softmax-reparameterization-for-output-head-quantization", "canonical_source": "https://aiflash.com/news/127850/", "published_at": "2026-09-28 14:30:00+00:00", "updated_at": "2026-09-28 14:48:57.678962+00:00", "lang": "en", "topics": ["large-language-models", "machine-learning", "ai-research", "ai-infrastructure"], "entities": [], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/softmax-reparameterization-for-output-head-quantization", "markdown": "https://wpnews.pro/news/softmax-reparameterization-for-output-head-quantization.md", "text": "https://wpnews.pro/news/softmax-reparameterization-for-output-head-quantization.txt", "jsonld": "https://wpnews.pro/news/softmax-reparameterization-for-output-head-quantization.jsonld"}}