论文多源确认精选

Stanford AI Lab 提出 hinted scaling laws,可从零分模型预测基准准确率

精选理由

Stanford AI Lab 把基准改成带提示的简单版本,小模型得零分也能算出 GSM8K 成绩,省 5600 倍算力,方法挺巧妙。

Stanford AI Lab 的 Suze van Adrichem 等人提出 hinted scaling laws 方法,把一个基准改造成带提示(hint)的多个更易版本,再联合建模模型规模 C 和提示等级 H 两个变量。这样即便小模型在原版基准上得零分,也能外推其准确率。实验显示,用该方法预测 GSM8K(H=0)的准确率,误差控制在 3 分以内,计算量减少 5600 倍。

原文 · Stanford AI Lab

Hinted scaling laws allow you to make better benchmark accuracy forecasts! Suze van Adrichem @suze_va Can we forecast benchmark accuracy from models that score zero? We (co-leads @LukeBailey181 @rckpudi ) turn a benchmark into a series of easier hinted versions. Jointly modeling scale C and hint level H enables forecasting GSM8K acc (H=0) within 3 pts from 5600x less compute 🧵 🔗 View Quoted Tweet 💬 0 🔄 2 ❤️ 3 👀 503 📊 1 ⚡