diff options
| author | Yuren Hao <yurenh2@illinois.edu> | 2026-08-03 17:47:13 -0500 |
|---|---|---|
| committer | Yuren Hao <yurenh2@illinois.edu> | 2026-08-03 17:47:13 -0500 |
| commit | 1e1d266db41f3acdcd647df71070982b2d1b057b (patch) | |
| tree | 59231947f311f51de67006e6bf6223d0edd57943 /ep_run/oracle_adjoint_train.py | |
| parent | 0918c65cc07b3d4a06be56a46f80a7c5f143471a (diff) | |
RESULT 88: K电池判决 — K=1/2/3在CE上不可分(配对差3e-5, 比seed散布小两个数量级), 成本可降至1.67×BP
三臂共种子=val批次相同=噪声完全相关, 配对差分为正统计量。先疑bit-identical(bf16空转)经逐行
比对否决: K有真实但可忽略的数值效应。诚实定位: K=1即BP+仅存于θ-雅可比的O(β)修正, 迭代在本
架构不做实质工作(级联自由相按构造为不动点, 与KHS同); 贡献在局部性/有限β偏差长程可控/硬件约束,
不在"梯度不同"。待决: 主线是否改K=1(倾向旗舰保K=3可比, 另开K=1成本臂)。
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
Diffstat (limited to 'ep_run/oracle_adjoint_train.py')
0 files changed, 0 insertions, 0 deletions