以下部分引用 AI 总结,现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
Delta Attention Residuals
4 atoms · 跨 1 天 · 首见 2026-05-20 · 最近 2026-05-20
三色: 🟦 fact 0 · 🟥 take 4 · stance ▲4/▼0/◆0
来源: X 4
时态: fresh:4
标签: 好数字:4
事实 (fact) (4)
- 2026-05-20 · sharper cross-layer routing
Delta Attention Residuals lift max attention weight from ~0.2 → ~0.6 (0.62 vs 0.35 avg) and curing routing collapse in deep layers.
tags: 好数字 · → daily
value: qty=max attention weight from ~0.2 → ~0.6 · avg=0.62 vs 0.35
- 2026-05-20 · validation PPL reduction at 7.6B scale
−𝟖.𝟐% 𝐯𝐚𝐥𝐢𝐝𝐚𝐭𝐢𝐨𝐧 𝐏𝐏𝐋 𝐚𝐭 𝟕.𝟔𝐁
tags: 好数字 · → daily
value: qty=-8.2% · scale=7.6B
- 2026-05-20 · parameter overhead
≤𝟎.𝟎𝟏% 𝐩𝐚𝐫𝐚𝐦𝐞𝐭𝐞𝐫 𝐨𝐯𝐞𝐫𝐡𝐞𝐚𝐝
tags: 好数字 · → daily
value: qty=≤0.01%
- 2026-05-20 · throughput compared to Attention Residuals
runs faster + lighter than Attention Residuals (14.0k vs 12.5k tok/s, 42.7 vs 44.0 GB)
tags: 好数字 · → daily
value: qty=14.0k vs 12.5k tok/s
时间轴 (近 20)
- 2026-05-20 ·
fact · sharper cross-layer routing · → daily
- 2026-05-20 ·
fact · validation PPL reduction at 7.6B scale · → daily
- 2026-05-20 ·
fact · parameter overhead · → daily
- 2026-05-20 ·
fact · throughput compared to Attention Residuals · → daily
← 实体目录 · 系统日志