site: surface K3 round six on homepage
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 04</span> ATTENTION RESIDUALS · PREREGISTERED PROBE</p>
|
||||
<h2>真实 K3 权重仍有冲突:先把一个可证伪的深度路由问题完整做完</h2>
|
||||
<p class="eyebrow"><span>NEW / K3 ROUND 06</span> SPIKE TRAJECTORY · SAME-FORWARD BACKWARD PATH</p>
|
||||
<h2>尖峰不是出生时就有:它在训练中形成,并对 value 路径敏感</h2>
|
||||
<p>
|
||||
不裁剪 <code>A_log [128]</code> 冒充 96-head K3 forward;冻结相同主干、初始化、数据窗口和预算,
|
||||
从零训练 Baseline / Full / Block 共 9 个 2,000-step 格。两种 AttnRes 的三 seed BPC
|
||||
配对方向都为负,但梯度均匀性指标没有复现论文叙述;支持与反证在同一实验室展示。
|
||||
严格复用 Round 05 的 depth-32 Block 正式格,追踪六个训练时点、六个张量位置与四种
|
||||
reduction。layer 21–25 的尖峰在 step 500 后形成;全局 uniform value-backward
|
||||
让 contrast 平均下降 70.2%,但只判为反向路径敏感性,不冒充因果贡献或训练变体。
|
||||
</p>
|
||||
</div>
|
||||
<dl>
|
||||
<div><dt>REPORT</dt><dd>16 Figures · 5 Tables</dd></div>
|
||||
<div><dt>GRID</dt><dd>3 structures × 3 seeds</dd></div>
|
||||
<div><dt>REPLAY</dt><dd>2,000 steps · 8/8 exact</dd></div>
|
||||
<div><dt>FORMAL</dt><dd>3 seeds × 8,000 steps</dd></div>
|
||||
<div><dt>ROBUST</dt><dd>4 reductions · 12/12</dd></div>
|
||||
<div><dt>REPLAY</dt><dd>16 frozen groups exact</dd></div>
|
||||
</dl>
|
||||
<span class="release-arrow" aria-hidden="true">进入训练曲线、深度权重与梯度反证 →</span>
|
||||
<span class="release-arrow" aria-hidden="true">进入训练轨迹、六位置谱与反向路径干预 →</span>
|
||||
</a>
|
||||
<a class="release-card deepseek-release" href="/deepseek/">
|
||||
<div>
|
||||
|
||||
Reference in New Issue
Block a user