{"type":"link","version":"1.0","title":"A trainable attention sink can be realized as a post-softmax output scale instead of an extra key-value pair","author_name":"AI Archs","author_url":"https://ai-arch.pages.dev","provider_name":"AI Archs","provider_url":"https://ai-arch.pages.dev","url":"https://ai-arch.pages.dev/n/attention-sink-as-output-scaling-not-extra-kv-pair","thumbnail_url":"https://ai-arch.pages.dev/og/attention-sink-as-output-scaling-not-extra-kv-pair.png","thumbnail_width":1200,"thumbnail_height":630}