SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Models

Jiaojiao Han1, Wujiang Xu2, Mingyu Jin2, Mengnan Du3,*

1 New Jersey Institute of Technology, 2 Rutgers University, 3 The Chinese University of Hong Kong, Shenzhen

* Corresponding author

Framework diagram for SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Models
Figure 1. SAGE tests and refines candidate feature explanations using activation feedback. Select the image to view it full size.

Overview

SAGE turns sparse autoencoder feature interpretation into an iterative process. Explainer agents propose possible descriptions, design targeted tests, inspect activation feedback, and refine their hypotheses. A final synthesis combines the supported explanations and evidence into a description of the feature.

Results

Table 2 of the paper compares SAGE with Neuronpedia on three language models. The selected layer-3 results below show higher generative and predictive accuracy for SAGE in each model.

ModelMethodGenerative accuracy ↑Predictive accuracy ↑
GPT-OSS-20BNeuronpedia0.260.62
GPT-OSS-20BSAGE0.590.80
Qwen3-4BNeuronpedia0.220.68
Qwen3-4BSAGE0.540.72
Gemma-2-2BNeuronpedia0.750.68
Gemma-2-2BSAGE0.970.83
Table 2. Selected layer-3 results; see the paper for layers 7, 11, and 23.

Qualitative Explanation Comparison

Table 3 shows how activation-guided tests narrow broad feature descriptions into more specific, testable explanations.

Model / featureNeuronpedia descriptionSAGE explanation
Gemma-2-2B · 11/5125Multithreading synchronization and thread safety.Code locking constructs in Python, C++, and Java; ordinary uses of “lock” do not trigger the feature.
Qwen3-4B · 23/24625English negative contractions involving “n’t”.A narrower detector for the “can’t” contraction suffix; tests with “won’t” and “don’t” refute the broader description.
GPT-OSS-20B · 3/121075Land-related contexts such as habitats and planets.Lexical or morpheme-level response to “terrestrial” and related fragments.
Gemma-2-9B-IT · 20/1Java exception and error-handling references.A lexical detector focused on “IllegalArgumentException” and related surface forms.
Table 3. Selected SAE feature explanations from the paper, summarized for readability.

BibTeX

@misc{han2026sage,
  title={SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Models},
  author={Han, Jiaojiao and Xu, Wujiang and Jin, Mingyu and Du, Mengnan},
  year={2026},
  eprint={2511.20820},
  archivePrefix={arXiv}
}