SabaPivot's picture
Rebuild canonical learning-rate balancing evidence
e51cce5 verified
Raw
History Blame Contribute Delete
867 Bytes
<!doctype html><html><head><meta charset="utf-8"><style>body{margin:0;background:#eef6f5;color:#17302d;font-family:Arial}.p{padding:26px;border:4px solid #20756c}.g{display:grid;grid-template-columns:repeat(3,1fr);gap:14px}.c{background:#fff;padding:16px;border-radius:12px}.b{font-size:29px;font-weight:800;color:#b45138}h1{margin:8px 0 18px}h2{font-size:17px;color:#20756c}</style></head><body><main class="p"><b>ICML 2026 REPRODUCTION</b><h1>Balancing layer learning rates</h1><section class="g"><div class="c"><h2>Closed-form losses</h2><div class="b">≤1.42 SEM</div><p>two- and three-layer MC agreement</p></div><div class="c"><h2>Transition exponents</h2><div class="b">1 and 2/3</div><p>exact symbolic Hessian balance</p></div><div class="c"><h2>Gradient identity</h2><div class="b">8.67e-18</div><p>maximum residual</p></div></section></main></body></html>