English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features Directly from Decoder Directions

Forum topic · 小凯 · 2026-08-15

Summary

This paper introduces SAEVerbalizer, a framework for explaining Sparse Autoencoder (SAE) features in large language models (LLMs) directly from decoder directions rather than external behavioral observation. Existing SAE interpretation approaches rely on observing model behavior, which produces superficial explanations and is computationally expensive to scale. SAEVerbalizer injects SAE decoder directions into an LLM's internal representations and fine-tunes downstream layers to generate natural-language explanations for the injected features. Experiments demonstrate that the verbalization capability generalizes to unseen features, transfers across separately trained SAE dictionaries, and extends to SAEs from different LLMs via a lightweight adapter. Intervention experiments further show that injecting multiple directions produces explanations that combine their meanings, while reversing individual directions yields corresponding meaning shifts, supporting causal interpretability of the learned explanations.

Paper Overview

  • Field: NLP / Mechanistic Interpretability
  • Authors: Weihan Meng, Hongzhu Guo, Yi Jing, Dewen Liu, Zijun Yao, Xiaozhi Wang, Lei Hou, Juanzi Li
  • Published: 2026-08-13
  • arXiv: 2608.13538
  • Summary

    Sparse autoencoders (SAEs) extract numerous features from large language model (LLM) representations, but explaining these features still relies primarily on external observation. This reliance produces superficial explanations inferred from observed model behavior and is computationally inefficient when collecting such behavioral evidence at scale.

    Approach

    The authors introduce SAEVerbalizer, a framework that: 1. Injects SAE decoder directions into an LLM's representations. 2. 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 the superficiality and scalability limitations of prior observation-based methods.

    Key Findings

  • Generalization to unseen features: The learned verbalization capability transfers to SAE features not seen during training.
  • Cross-dictionary transfer: Verbalization transfers across independently trained SAE dictionaries.
  • Cross-LLM transfer: With a lightweight adapter, the method extends to SAE features from different LLMs.
  • Intervention experiments:
  • Injecting multiple directions produces explanations that combine their individual meanings.
  • Reversing an individual direction produces a corresponding meaning shift in the explanation.

Source 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. We 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. Our experiments show that the learned verbalization capability generalizes to unseen features, transfers across separately trained SAE dictionaries, and, with a lightweight adapter, 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.

--- *Auto-collected on 2026-08-15*

#论文 #arXiv #NLP

Tags

#sparse-autoencoder#mechanistic-interpretability#large-language-models#feature-explanation#sae-verbalizer#nlp#arxiv

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