Summary
SAEVerbalizer is a new framework for interpreting sparse autoencoder (SAE) features in large language models (LLMs), presented in arXiv paper 2608.13538 by Weihan Meng and colleagues. Traditional SAE feature interpretation relies on external observation of model behavior, which produces superficial explanations and is computationally inefficient at scale. SAEVerbalizer instead injects SAE decoder directions into an LLM's representations and fine-tunes the model's downstream layers to generate natural-language explanations of the injected features. Once trained, the verbalizer explains SAE features directly from decoder directions, addressing both limitations. Experiments demonstrate that the learned verbalization capability generalizes to unseen features, transfers across independently trained SAE dictionaries, and extends to SAE features from different LLMs via a lightweight adapter. Intervention experiments show that injecting multiple directions yields explanations combining their meanings, while reversing individual directions produces corresponding meaning shifts in the generated explanations.
Paper Overview
Research Area: NLP
Authors: Weihan Meng, Hongzhu Guo, Yi Jing, Dewen Liu, Zijun Yao, Xiaozhi Wang, Lei Hou, Juanzi Li
arXiv: 2608.13538
Abstract
Sparse autoencoders (SAEs) are proposed to extract numerous features from large language model (LLM) representations, yet explaining these features still relies primarily on external observation. This reliance leads to superficial explanations inferred from observed model behavior and computational inefficiency from collecting such behavioral evidence at scale. The authors introduce SAEVerbalizer, a framework that injects SAE decoder directions into an LLM's representations and fine-tunes the LLM's downstream layers to generate natural-language explanations of the injected features.
Once trained, the resulting verbalizer explains SAE features directly from decoder directions, addressing both limitations.
Key Findings
- The learned verbalization capability generalizes to unseen features.
- It transfers across separately trained SAE dictionaries.
- With a lightweight adapter, it extends to SAE features from different LLMs.
- Intervention experiments show that injecting multiple directions yields an explanation combining their meanings, while reversing individual directions produces corresponding meaning shifts.
Links
- Paper: https://arxiv.org/abs/2608.13538
*Auto-collected on 2026-08-15*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178633504