推荐yurimo · 2026-11
“a paper I particularly like: Tracr”
推荐这篇自己特别喜欢的编译式 Transformer 前人工作
把可读程序编译成 Transformer 权重,为可解释性研究提供已知真值。
提出 Tracr 编译器,把人类可读的程序编译成标准的仅解码器(decoder-only)Transformer 模型,使模型结构完全已知。这种已知结构可用来设计实验(例如研究“叠加”superposition),也可作为评估可解释性方法的真值(ground truth)。论文用统计词频、排序、括号检查等程序做了演示,并开源了实现。
评论者把它作为“把程序或代码直接编译进 Transformer 权重、无需训练”这一方向的相关工作来引用。
“a paper I particularly like: Tracr”
“Easy to say obviously but there's some promising work in this direction[1]”
“Follow on work actually demonstrated how RASP-like programs could actually be compiled into model weights without training [3].”
“See this paper from DeepMind which talks about compiling existing code into transformer primitives Tracr”
“Tracr: Compiled Transformers as a Laboratory for Interpretability Tracr (code at [LINK])”