research: publish AttnRes forward training study
This commit is contained in:
@@ -128,20 +128,20 @@ const paths = [
|
||||
<div class="release-grid">
|
||||
<a class="release-card k3-release" href="/k3/">
|
||||
<div>
|
||||
<p class="eyebrow"><span>NEW / K3 ROUND 07</span> LOCAL MIXER PATH · BIDIRECTIONAL GATE</p>
|
||||
<h2>16 个局部 mixer 单侧证据很强,但双向定位仍然没有闭合</h2>
|
||||
<p class="eyebrow"><span>NEW / K3 ROUND 08</span> TRAIN-TIME FORWARD · PREREGISTERED GATE</p>
|
||||
<h2>局部 uniform routing 进入完整训练后,六个尖峰指标格全部衰减</h2>
|
||||
<p>
|
||||
以 14 个冻结 mask 同时检查 sufficiency 与 restoration:groups 6+7 在 learned
|
||||
背景的 6 / 6 格全部超过 50% global log gap;但从 uniform 背景恢复时,peak 三个
|
||||
seed 全部未过线。因此只能报告 one-sided evidence,不能宣布尖峰已定位到这 16 个 mixer。
|
||||
四个前向架构变体各跑三个 8,000-step seed,再完整 replay 主格:groups 6+7 的
|
||||
contrast / peak 六格降幅为 32.3%–77.3%,BPC 质量门 4 / 4 通过。结论只限缩小
|
||||
depth-32 Block 协议,不是真实 K3 checkpoint 或 Figure 5(c) 复现。
|
||||
</p>
|
||||
</div>
|
||||
<dl>
|
||||
<div><dt>MATRIX</dt><dd>14 masks × 3 seeds</dd></div>
|
||||
<div><dt>SUFFICIENCY</dt><dd>6 / 6 pass</dd></div>
|
||||
<div><dt>RESTORATION</dt><dd>3 / 6 fail</dd></div>
|
||||
<div><dt>MATRIX</dt><dd>12 formal + 1 replay</dd></div>
|
||||
<div><dt>ATTENUATION</dt><dd>6 / 6 pass</dd></div>
|
||||
<div><dt>QUALITY</dt><dd>4 / 4 pass</dd></div>
|
||||
</dl>
|
||||
<span class="release-arrow" aria-hidden="true">进入局部路径图、双向门与 32 层原始谱 →</span>
|
||||
<span class="release-arrow" aria-hidden="true">进入训练轨迹、主门、non-additivity 与 32 层谱 →</span>
|
||||
</a>
|
||||
<a class="release-card deepseek-release" href="/deepseek/">
|
||||
<div>
|
||||
|
||||
Reference in New Issue
Block a user