跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2607.09800 · v5

Predicting Lost Updates Is Not Enough to Choose a Rounding Policy

首次提交 2026-07-09 23:23
版本更新 2026-11-18 14:09
雷达观测 2026-11-19 19:08
arXiv ↗PDF ↗

摘要

In low-precision training, a weight can round back to its previous value after a small update. We test whether predicting how often this happens helps choose a rounding rule. In Pythia-410M continuations from a common checkpoint, a higher-precision pilot accurately predicted the fraction of matrix weights left unchanged at each of the first 64 updates. Across three schedules with optimizer state retained or reset, nearest rounding gave lower validation loss at 256 updates than stochastic rounding or retaining 32-bit weights. The same ordering held at 512 updates with retained state. Accurately forecasting the unchanged fraction did not identify a benefit from switching policies in these runs.

作者备注:13 pages, 3 figures, 8 tables. This version adds matched Pythia-410M continuations under three weight-update policies, a paired optimizer-state comparison, and a 512-update extension. Ancillary files provide numerical data and replay scripts for all 36 runs

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v5,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v5 · 2026-11-18 14:09 · 最新观测版本

Predicting Lost Updates Is Not Enough to Choose a Rounding Policy

In low-precision training, a weight can round back to its previous value after a small update. We test whether predicting how often this happens helps choose a rounding rule. In Pythia-410M continuations from a common checkpoint, a higher-precision pilot accurately predicted the fraction of matrix weights left unchanged at each of the first 64 updates. Across three schedules with optimizer state retained or reset, nearest rounding gave lower validation loss at 256 updates than stochastic rounding or retaining 32-bit weights. The same ordering held at 512 updates with retained state. Accurately forecasting the unchanged fraction did not identify a benefit from switching policies in these runs.

查看此版本解读 →查看此版本原文 ↗