Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders

Dong Shu, Xuansheng Wu, Haiyan Zhao, Mengnan Du, Ninghao Liu

Abstract

Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs). However, conventional approaches to analyzing SAEs typically rely solely on input-side activations, without considering the causal influence between each latent feature and the model's output. This work is built on two key hypotheses: (1) activated latents do not contribute equally to the construction of the model's output, and (2) only latents with high causal influence are effective for model steering. To validate these hypotheses, we propose Gradient Sparse Autoencoder (GradSAE), a simple yet effective method that identifies the most influential latents by incorporating output-side gradient information.

Introduction

Mechanistic interpretability aims to decode the internal "black box" of Large Language Models (LLMs). Sparse Autoencoders (SAEs) are currently a leading method for this, decomposing dense representations into sparse, human-interpretable features or "latents." A common assumption in the field is that if a latent is highly activated by an input, it must be significant for the model's output.

However, this research challenges that assumption, introducing the concept of "bystander" latents—features that fire strongly during processing but have little to no causal impact on the final output. To address this, we introduce GradSAE, a method that uses gradient information to identify which latents actually drive model behavior.

Methodology: The GradSAE Framework

The core innovation of GradSAE is moving beyond simple activation magnitude. Instead of assuming importance based on activation $H$, GradSAE calculates an influence score $g$ by considering the gradient of the model's output with respect to that latent.

The GradSAE framework identifies influential latents by combining activations with output-side gradients.

Ideally, measuring influence would involve ablating every latent to observe output changes, but this is computationally prohibitive. GradSAE approximates this causal influence using a first-order Taylor expansion. The influence score for a latent is the element-wise product of its activation and its gradient: $$g_{n,c} \approx \frac{\partial p(Y|h(Z))}{\partial H_{n,c}} \odot H_{n,c}$$ This approach is training-free and can be applied to existing SAEs and LLMs as a simple inference-time addition.

Experimental Setup

The researchers evaluated GradSAE using the SQuAD (Stanford Question Answering Dataset) on models like Gemma 2 9B (using Gemma Scope SAEs) and LLaMA 3 8B. SQuAD is particularly useful because it allows for "Local Steering" experiments: attempting to steer the model toward an answer for one question while it processes a context shared with another question.

Key Results and Discussion

The experiments revealed a striking "Bystander Effect": many highly active latents are essentially irrelevant to the task. The overlap between the most active and most influential latents was found to be only 18%.

Experimental results comparing GradSAE against baseline activation-based methods in perturbation and steering tasks.

Perturbation Analysis

In perturbation tests, masking the top 50% of influential latents identified by GradSAE caused the model's Exact Match (EM) score to drop from 100% to 30.58%. Conversely, masking the top 50% based on activation magnitude only reduced the score to 53.42%. This demonstrates that GradSAE is significantly more accurate at identifying the "load-bearing" features of a representation.

Model Steering

GradSAE also proved superior for model steering. In the local steering task, GradSAE achieved a steering F1 score of 10.21% (at K=10), outperforming the activation-based baseline of 6.69%. This confirms that interventions are most effective when targeted at latents with high causal influence rather than just high activation.

Conclusion

GradSAE provides a more nuanced and causally grounded approach to LLM interpretability. By demonstrating that activation does not equal influence, this work highlights the necessity of considering output-side gradients when interpreting or steering model representations. Future work will explore applying GradSAE to identify latents responsible for safety risks and social biases, where precise intervention is paramount.

Summary updated

Abstract

Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs).However, conventional approaches to analyzing SAEs typically rely solely on input-side activations, without considering the causal influence between each latent feature and the model's output.This work is built on two key hypotheses: (1) activated latents do not contribute equally to the construction of the model's output, and (2) only latents with high causal influence are effective for model steering.To validate these hypotheses, we propose Gradient Sparse Autoencoder (GradSAE), a simple yet effective method that identifies the most influential latents by incorporating output-side gradient information.
BibTeX
@misc{shu2025beyond,
  author = {Dong Shu and Xuansheng Wu and Haiyan Zhao and Mengnan Du and Ninghao Liu},
  title = {{Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders}},
  year = {2025},
  eprint = {2505.08080},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.08080},
}

Loading related content...

Figure