---
title: "When Are Two SAE Latents the Same Feature?"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2611.05655v1"
permalink: "https://feed.astralor.com/items/itm_2ec62f25eb8c647ad120f3f6"
published: "2026-11-03T21:39:53.000Z"
score: 55
channel: "tech"
tags: ["推理效率","When Are Two","SAE Latents","Same Feature","Comparing","SAEs"]
---

# When Are Two SAE Latents the Same Feature?

## 速览

Comparing sparse autoencoders (SAEs) requires connecting feature correspondence, differences between whole dictionaries, and uncertainty from finite evaluation data. We develop a framework that connects these tasks thro…

## 关键点

- Comparing sparse autoencoders (SAEs) requires connecting feature correspondence, differences between whole dictionaries, and uncertainty from finite evaluation data.
- We develop a framework that connects these tasks through an energy-weighted optimal-transport plan under a specified similarity geometry.
- The same plan defines a dictionary-level discrepancy, decomposes it into latent contributions, and quantifies how each latent's mass is distributed across counterparts.

## 全文

Comparing sparse autoencoders (SAEs) requires connecting feature correspondence, differences between whole dictionaries, and uncertainty from finite evaluation data. We develop a framework that connects these tasks through an energy-weighted optimal-transport plan under a specified similarity geometry. The same plan defines a dictionary-level discrepancy, decomposes it into latent contributions, and quantifies how each latent's mass is distributed across counterparts. We characterize relationships between decoder, activation, and first-order causal similarities; idealized models clarify the interpretation of splitting and absorption. For fixed trained SAEs, we derive joint asymptotic inference for transport costs and fractions, accounting for activation similarities and both energy marginals estimated from the same documents. A joint document bootstrap gives asymptotically valid simultaneous confidence bands. Experiments across four language models locate cross-seed discrepancy and separate context-dependent changes.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_2ec62f25eb8c647ad120f3f6