跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.12206 · v1

A Geometric Invariant of Attention Predicts Head Specialization from Weights Alone

首次提交 2026-11-09 19:02
版本更新 2026-11-09 19:02
雷达观测 2026-11-10 15:22
arXiv ↗PDF ↗

摘要

Transformers are built with continuous symmetries in the attention mechanism. We derive ρρ, a scalar invariant that measures each head's capacity to couple to rotary positional encodings. This framework allows us to analyze heads without forward passes, opening a new research direction into data-independent analysis of modern LLMs. Empirically, we find that ρρ predicts induction heads, the most-studied circuit type in mechanistic interpretability, with Spearman correlation between −0.55-0.55 and −0.74-0.74 across four bias-free RoPE models spanning 135135M to 77B parameters. As a preliminary proof of concept, we use ρρ to apply YaRN to only the top-50% positional heads, which degrades perplexity less than uniform application. Code is available at https://github.com/MPedraBento/rho-invariant.

作者备注:Accepted as an extended abstract at the Symmetry and Geometry in Neural Representations Workshop (NeurReps), NeurIPS 2026

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

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

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。

v1 · 2026-11-09 19:02 · 最新观测版本

A Geometric Invariant of Attention Predicts Head Specialization from Weights Alone

Transformers are built with continuous symmetries in the attention mechanism. We derive ρρ, a scalar invariant that measures each head's capacity to couple to rotary positional encodings. This framework allows us to analyze heads without forward passes, opening a new research direction into data-independent analysis of modern LLMs. Empirically, we find that ρρ predicts induction heads, the most-studied circuit type in mechanistic interpretability, with Spearman correlation between −0.55-0.55 and −0.74-0.74 across four bias-free RoPE models spanning 135135M to 77B parameters. As a preliminary proof of concept, we use ρρ to apply YaRN to only the top-50% positional heads, which degrades perplexity less than uniform application. Code is available at https://github.com/MPedraBento/rho-invariant.

查看此版本解读 →查看此版本原文 ↗