Enhancing SAE-based Steering via Neighbor Integrated Feature Selection

Yutian Liu; Xu Wang; Difan Zou

cs.AIAug 28, 2026

Enhancing SAE-based Steering via Neighbor Integrated Feature Selection

Authors: Yutian Liu, Xu Wang, Difan Zou

Organizations: The University of Hong Kong · University of Science and Technology of China

Abstract

Sparse autoencoders (SAEs) disentangle model activations into interpretable features and are widely used for steering large language models. Most existing SAE-based steering methods select features by applying a top- filter based on statistical scores, assuming that higher-scoring features yield stronger steering effects. In this paper, we show that this assumption is often invalid, leading to suboptimal feature selection. Our analysis reveals that effective steering features may be distributed among representationally adjacent, semantically similar groups induced by feature splitting in SAEs. Within such groups, features may exhibit disparate statistical scores despite having comparable steering influence, causing score-based selection to overlook important features. Based on these observations, we propose Neighbor Integrated Feature Selection (NIFS), a plug-and-play strategy that leverages representation similarity to improve feature selection for steering. We evaluate NIFS across multiple SAE-based steering methods and tasks, and demonstrate consistent performance gains over conventional top- $k$ selection.

Figures & tables

Panel of Figure 1 — Figure 1: Steering success rate on Gemma2-2B across different top- $k$ values ranging from 100 to 1500. The performance is reported for three tasks: (a) Knowledge Conflicts, (b) Sentiment, and (c) Detoxification.

Figure 2: Distribution of success rate over 32 runs of random feature selection on Gemma2-2B.

Figure 3: Cosine similarity between SAE features and probing directions across feature ranks for the Sentiment task on Gemma2-2B. Features are ranked by their importance scores under different methods. Blue points correspond to individual features, while the red line indicates an exponential fit trend. Negative and positive probing directions are shown in separate rows.

Model	Method	KC	Sentiment		Detoxification
Model	Method	SR	SR	Flu.	SR	Flu.
Llama2-7B	Vanilla	$38.14{\scriptstyle\pm 3.97}$	$38.16{\scriptstyle\pm 1.31}$	$4.43{\scriptstyle\pm 0.18}$	$62.79{\scriptstyle\pm 0.81}$	$3.92{\scriptstyle\pm 0.28}$
	CAA	$63.61{\scriptstyle\pm 3.56}$	$44.95{\scriptstyle\pm 0.99}$	$4.37{\scriptstyle\pm 0.13}$	$67.94{\scriptstyle\pm 1.00}$	$3.92{\scriptstyle\pm 0.27}$
	Probe	$65.38{\scriptstyle\pm 1.55}$	$47.56{\scriptstyle\pm 1.18}$	$4.30{\scriptstyle\pm 0.04}$	$71.40{\scriptstyle\pm 0.90}$	$3.80{\scriptstyle\pm 0.01}$
	Diffmean	$68.49{\scriptstyle\pm 3.04}$	$59.67{\scriptstyle\pm 0.95}$	$3.98{\scriptstyle\pm 0.17}$	$80.80{\scriptstyle\pm 1.06}$	$3.83{\scriptstyle\pm 0.01}$
	+NIFS	$\textbf{68.57}{\scriptstyle\pm 0.88}$	$\textbf{62.69*}{\scriptstyle\pm 1.29}$	$3.95{\scriptstyle\pm 0.04}$	$\textbf{85.30*}{\scriptstyle\pm 0.86}$	$3.84{\scriptstyle\pm 0.01}$

Table 1: Comparison of original and NIFS-enhanced SAE-based baselines across models and tasks. SR denotes the success rate and Flu. denotes the n-gram fluency.

Panel of Figure 4 — Figure 4: Steering success rate (SR) and average cosine similarity of neighbor features on Gemma2-2B for the Detoxification task across different neighbor sizes (top- $k$ ). The four subfigures correspond to the four methods: SPARE, SAIF, STA, and Diffmean.

Methods	KC	Detox		Sentiment
	SR	SR	Flu.	SR	Flu.
Diffmean	63.03	70.98	3.957	60.82	4.065
+NIFS	63.98	71.23	3.963	63.57	4.047
SAIF	62.53	73.64	3.948	59.21	4.258
+NIFS	62.78	73.89	3.938	60.46	4.203
STA	62.68	70.89	3.948	56.73	4.271

Table 2: Results on Gemma3-12B. SR denotes the success rate and Flu. denotes the n-gram fluency.

Method	Diffmean	SAIF	STA	SPARE
Vanilla	68.57	66.46	69.10	70.42
Only-core	65.64	64.55	64.13	68.56
Core-random	66.97	65.63	64.05	69.06
Larger- $k$	66.21	64.84	65.79	69.06
NIFS	70.22	67.65	70.14	71.64

Table 3: Comparison between NIFS and other simple feature selection strategies on the detoxification task using Gemma2-2B.

Methods	Diffmean	SAIF	STA	SPARE
Baseline	68.49	68.93	69.97	69.71
NIFS w/o FR	68.73	70.72	70.23	71.34
NIFS	68.57	70.79	71.16	71.56

Table 4: Ablation results for detoxification on Gemma2-2B. Each cell shows SR (%) for the corresponding method and setting.

Rank	Role	Explanation of Feature	Feature ID	Cos sim
1	Core	references to creativity and personal expression related to arts and design	48234	-
	Neighbor	emoticons and expressions of emotion	12757	0.222
	Neighbor	expressions of surprise or disbelief	50409	0.179
2	Core	discussions about differing perspectives and values in relationships	36077	-
	Neighbor	references to women’s health issues in the workplace	20722	0.536
	Neighbor	technical errors and conditions related to coding or programming	55072	0.441

Table 5: Top-5 SAE features selected by Diffmean and their neighbor features identified by NIFS for the Detoxification task on Gemma2-2B. Feature explanations are retrieved from Neuronpedia Lin (2023) .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

You are a sentiment evaluator. Your task is to determine whether the Steered continuation successfully transforms the sentiment of the Original Input from negative or neutral to positive. Rate the output from 0 to 5 based on the emotional tone of the continuation. Don’t be influenced by the sentiment of original prompt.

- 1 indicates a completely negative sentiment.

- 3 indicates a well-balanced or neutral sentiment.

- 5 indicates a completely positive sentiment.

Please respond in the following format:

score: x out of 5

Appendix

Table 6: Evaluation Prompt for Sentiment Task Steering Assessment

Panel of Figure 5 — Figure 5: Steering success rate on Llama2-7B across different top- $k$ values ranging from 100 to 1500.

Figure 6: Distribution of success rate over 32 runs of random feature selection on Llama2-7B.

Figure 7: Cosine similarity between SAE features and probing directions across feature ranks for the Knowledge Conflicts task on Gemma2-2B.

Figure 8: Cosine similarity between SAE features and probing directions across feature ranks for the Detoxification task on Gemma2-2B.

Figure 9: Cosine similarity between SAE features and probing directions across feature ranks for the Knowledge Conflicts task on Llama2-7B.

Figure 10: Cosine similarity between SAE features and probing directions across feature ranks for the Detoxification task on Llama2-7B.

Figure 11: Cosine similarity between SAE features and probing directions across feature ranks for the Sentiment task on Llama2-7B.

Methods	Detox	KC	Sentiment
Diffmean	$68.49{\scriptstyle\pm 3.04}$	$80.80{\scriptstyle\pm 1.06}$	$59.67{\scriptstyle\pm 0.95}$
NIFS w/o FR	$\textbf{68.73}{\scriptstyle\pm 1.13}$	$83.78{\scriptstyle\pm 3.30}$	$60.35{\scriptstyle\pm 2.41}$
+NIFS	$68.57{\scriptstyle\pm 0.88}$	$\textbf{85.30}{\scriptstyle\pm 0.86}$	$\textbf{62.69}{\scriptstyle\pm 1.29}$
SAIF	$68.93{\scriptstyle\pm 3.32}$	$73.10{\scriptstyle\pm 0.49}$	$47.42{\scriptstyle\pm 1.24}$
NIFS w/o FR	$70.72{\scriptstyle\pm 1.69}$	$73.22{\scriptstyle\pm 1.10}$	$\textbf{48.97}{\scriptstyle\pm 0.53}$
+NIFS	$\textbf{70.79}{\scriptstyle\pm 1.58}$	$\textbf{74.63}{\scriptstyle\pm 1.00}$	$48.72{\scriptstyle\pm 0.71}$

Appendix

Table 7: Ablation study of feature reweighting across different tasks on Gemma2-2B.

Language Model Steering

Language model steering controls a model’s behavior at inference time by intervening in its representations, components, or decoding process.

See all 281 papers

Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects
Jun 6, 2026Evan DuanLanguage Model Steering LLM Interpretability
Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselines
May 29, 2026Mikkel Godsk Jørgensen, Lars Kai HansenLanguage Model Steering LLM Interpretability
DTU Compute · Technical University of Denmark
Look Before You Steer: Geometry Predicts SAE Feature Steerability
Sep 19, 2026Muhammad Khan, Shlok Channawar, Akshaj Gurugubelli +2Sparse Autoencoders Activation Steering
Algoverse AI Research.

Jun 6, 2026J/K move · Enter open · S save

Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects

Evan Duan

Language Model Steering LLM Interpretability Sparse Autoencoders Causal Interventions in Language Models

Open paper AskPDF arXiv

Sparse autoencoder (SAE) features are increasingly used to steer language models, but feature steering is rarely clean: the same intervention can behave inconsistently across contexts and perturb unrelated features. We introduce a pre-intervention screening framework for forecasting SAE steering side effects from feature statistics computed before steering. We operationalize side effects along two axes of steering modularity, effect stability and collateral spread, and evaluate GPT-2-small, Pythia-70M-deduped, Gemma-2-2B, and Llama-3.1-8B across ReLU, JumpReLU, and TopK SAE dictionaries. Across these settings, decoder geometry, activation statistics, co-activation structure, and direct-logit footprint predict steering modularity better than frequency-only and activation-magnitude baselines. The signal is strongest in GPT-2-small, Pythia-70M, and Llama-3.1-8B, where it survives residualization against magnitude-related confounds, and weaker in Gemma-2-2B. Held-out screening shows that ranking unseen features by predicted cleanliness can select features that steer more cleanly on fresh contexts, but the successful axis varies by setting: GPT-2 improves most cleanly, Pythia improves mainly on stability, Llama mainly on collateral, and Gemma only partially. A controlled Llama Scope width comparison shows that the predictive signal persists under a 32K-to-128K dictionary-width change, although the screening payoff becomes less stable. Overall, SAE steering side effects are predictable in advance, but the useful predictor signature and transferred modularity axis are model- and dictionary-setting dependent.

Abstract

Language Model Steering

Explore similar work

Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects

Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselines

Look Before You Steer: Geometry Predicts SAE Feature Steerability