PaperlayerView exact revision on arXiv

Source-linked AI summary

Enhancing LLM Steering through Sparse Autoencoder-Based Vector Refinement

Anyi Wang, Xuansheng Wu, Dong Shu, Yunpu Ma, Ninghao Liu

arXiv:2509.23799v2cs.LGcs.AI

TL;DR

Steering vectors learned from small datasets can contain task-irrelevant features, and common SAE feature selection can still pick superficial ones. SAE-RSV uses feature semantics to denoise and augment those vectors, with experiments reporting consistent gains over baselines including fine-tuning.

  • Problem

    Steering methods often rely on large datasets, while vectors learned from small samples can be noisy and prior top-k SAE selection may choose superficial task-irrelevant features.

    Source: §1 · pp. 1–2

  • Method

    SAE-RSV uses an LLM to judge feature semantics, removes task-irrelevant features, and adds missing relevant features to steering vectors learned from limited data.

    Source: §1 · p. 2

  • Results

    Across five concept datasets with 50 training sample pairs, SAE-RSV consistently outperforms all baselines, including fine-tuning.

    Source: §1 · p. 2, §6 · p. 10

  • Takeaways & Limitations

    The findings indicate that steering vectors can be constructed from limited training data by refining them with SAE-based denoising and augmentation.

    Source: §6 · pp. 9–10

    Prior SAE feature selection can still select superficial task-irrelevant features such as punctuation or stop words.

    Source: §1 · p. 2

Abstract

from arXiv · show

Steering has emerged as a promising approach in controlling large language models (LLMs) without modifying model parameters. However, most existing steering methods rely on large-scale datasets to learn clear behavioral information, which limits their applicability in many real-world scenarios. The steering vectors extracted from small dataset often contain task-irrelevant noising features, which degrades their effectiveness. To refine the steering vectors learned from limited data, we introduce Refinement of Steering Vector via Sparse Autoencoder (SAE-RSV) that leverages SAEs to semantically denoise and augment the steering vectors. In our framework, we first remove task-irrelevant features according to their semantics provided by SAEs, and then enrich task-relevant features missing from the small dataset through their semantic similarity to the identified relevant features. Extensive experiments demonstrate that the proposed SAE-RSV substantially outperforms all the baseline methods including supervised fine-tuning. Our findings show that effective steering vector can be constructed from limited training data by refining the original steering vector through SAEs.

1 INTRODUCTION

Steering can improve LLM controllability without retraining, but established approaches often rely on large datasets, limiting their use when only a small number of samples are available. SAE-RSV uses feature semantics to remove irrelevant features and recover missing relevant ones; experiments report gains over baselines including fine-tuning.

1 INTRODUCTION

  • Steering modifies LLM internal representations to guide behavior without retraining and has been explored for alignment, reasoning, and safety.The introduction describes steering as an alternative to fine-tuning and prompting.Source: §1 · p. 1
  • Steering vectors learned from small datasets can contain noise from spurious correlations, and prior top-k SAE feature selection may choose superficial features such as punctuation or stop words.This motivates data-efficient steering that preserves effectiveness and interpretability with limited data.Source: §1 · pp. 1–2
  • SAE-RSV semantically identifies task-relevant features, removes irrelevant features, and adds missing relevant features to refine steering vectors learned from limited data.The method uses an LLM to judge feature relevance, then retrieves missing features semantically similar to relevant ones and distinct from irrelevant ones.Source: §1 · p. 2
  • Across five datasets with 50 training sample pairs, SAE-RSV consistently outperforms all baselines, including fine-tuning.The analysis also finds that over 93.6% of features in original steering vectors are noisy, while those vectors capture only 42.2% of task-relevant features.Source: §1 · p. 2

2 PRELIMINARY: SPARSE AUTOENCODERS

Sparse autoencoders map dense LLM activations into sparse, interpretable latent features and reconstruct the input from active features. Their overcomplete latent representation activates only a small subset of features for each input.

2 PRELIMINARY: SPARSE AUTOENCODERS

  • An SAE encodes dense residual-stream activations into sparse latent features and decodes active features to reconstruct the original activation.The latent features are described as interpretable components of dense LLM representations.Source: §2 · p. 3
  • The latent dimension is typically larger than the input dimension, but only a small subset of components is active for each input.This yields a highly sparse representation.Source: §2 · p. 3
  • The decoder linearly combines active latent features to reconstruct the original input, while the feature activations remain sparse.Each decoder row corresponds to a learned feature vector, and nonzero activations indicate which features are active.Source: §2 · p. 3

3 PROPOSED SAE-RSV APPROACH

SAE-RSV refines limited-data steering by identifying task-irrelevant features, retrieving missing relevant features, and combining both corrections with the original steering vector.

3.1 NOISES IN STEERING VECTORS

  • The conventional steering vector averages positive-minus-negative hidden-representation differences, but limited and incomplete training data can introduce noise and coverage bias.Source: §3.1 · p. 3

3.2 NOISE VECTOR CONSTRUCTION

  • SAE-RSV seeds candidates from feature activation differences, then uses an LLM to judge semantic relevance and separate noisy from task-relevant features.Source: §3.2 · p. 4

3.3 USEFUL VECTOR CONSTRUCTION

  • To recover uncaptured relevant features, SAE-RSV ranks features by explanation similarity to relevant rather than noisy features, then verifies and averages selected features.Verification can be manual or LLM-based; the selected set comprises the top-K features.Source: §3.3 · pp. 4–5

3.4 STEERING VECTOR DENOISING AND AUGMENTATION

  • The final vector combines the original steering direction with a subtracted noise vector and an added useful vector, balancing fidelity, denoising, and enrichment.At inference, the refined vector is injected into hidden representations at each token position.Source: §3.4 · p. 5

4 EXPERIMENTS

The experiments compare SAE-RSV with steering baselines and fine-tuning, then examine its modules, feature counts, hyperparameters, data-size sensitivity, and feature semantics.

4.1 EXPERIMENTAL SETTINGS

  • Experiments use Llama-3-8B-Instruct, five behavioral datasets, and 50 training pairs per concept, comparing SAE-RSV with seven baselines using success rate and entropy.SAE-RSV feature relevance is judged by GPT-4o-mini; the baselines include LoRA-SFT and CAA.Source: §4.1 · pp. 5–6

4.2 EFFECTIVENESS FOR MODEL STEERING

  • Across five behavioral concepts, SAE-RSV beats all baselines, averages 18.8% higher success than the original prompt, and exceeds CAA by over 10% on four datasets.It also achieves substantially higher success rates than LoRA-SFT across tasks; the authors report stronger out-of-distribution generalization than supervised fine-tuning.Source: Table 1 · p. 6, §4.2 · p. 6

4.3 EFFECTIVENESS COMPARISON OF DIFFERENT MODULES

  • Across five datasets, augmentation raises success from 45.6% to 52.8%, while denoising raises it from 45.6% to 49.2%.Denoising reduces entropy by 0.11, from 7.13 to 7.02; augmentation increases it by 0.11, from 7.13 to 7.24.Source: §4.3 · p. 7, Table 2 · p. 7

4.4 INFLUENCE OF FEATURE COUNT ON STEERING PERFORMANCE

  • Among positively activated features, about 104 are task-irrelevant on average; steering peaks around 15–20 features and surpasses CAA from 13 features.CAA captures less than half of the useful features; performance declines after SAE-RSV’s peak feature range.Source: §4.4 · pp. 7–8, Figure 2 · p. 7

4.5 INFLUENCE OF DIFFERENT HYPERPARAMETERS

  • On Corrigibility, increasing α2 and α3 can improve success, but excessively large values reduce it; performance remains relatively stable across scaling factors.When both α2 and α3 are set to 20, success rate drops to 90%.Source: Table 4 · p. 8, §4.5 · p. 8

4.6 SENSITIVITY ANALYSIS UNDER DIFFERENT TRAINING DATA SIZES

  • Across training sizes on Sycophancy, SAE-RSV consistently beats baselines and maintains a stable relative improvement over them.With 10 training samples, it matches CAA trained on 1000 samples and surpasses fine-tuning at the same training size.Source: §4.6 · p. 8

4.7 CASE STUDY

  • In Myopic Reward, identified relevant features commonly concern urgency, while numeric values and comparisons are treated as noise; example steering preserves fluency and structure.The authors present these feature selections as supporting semantic selection over selection based purely on dataset statistics.Source: §4.7 · pp. 8–9, Table 5 · p. 9

5 RELATED WORK

Prior steering-vector work uses Difference in Means (DoM), while SAE methods identify task-relevant features through activation differences. Such feature selection can retain task-irrelevant features, motivating more precise selection.

5 RELATED WORK

  • DoM constructs steering vectors by averaging activation differences between contrastive prompt sets.Prior studies report that DoM can capture task-relevant directions for reasoning and alignment interventions without fine-tuning.Source: §5 · p. 9
  • SAEs project dense activations into sparse latent spaces that yield human-interpretable features used to explain model behavior.Recent methods compare SAE feature activations across contrastive sample pairs to detect task-relevant features.Source: §5 · p. 9
  • Activation-difference selection can include task-irrelevant features such as punctuation or stop words, motivating more precise feature selection.Prior work also reports noise in DoM vectors from training data.Source: §5 · p. 9

6 CONCLUSION

SAE-RSV refines steering vectors through Denoising and Augmentation, discarding topic-irrelevant features and adding inactivated topic-relevant ones. Experiments report improved steering across five concept datasets, alongside robustness and generation-quality findings.

6 CONCLUSION

  • Denoising discards topic-irrelevant features, while Augmentation adds inactivated topic-relevant features to the original steering vector.The method's performance is reported as robust to hyperparameter combinations and training data sizes.Source: §6 · pp. 9–10
  • SAE-RSV consistently outperforms other baselines without compromising generation quality on Llama-3-8B-Instruct across five concept datasets with 50 training sample pairs.Both Denoising and Augmentation contribute to improved steering performance.Source: §6 · p. 10
  • Topic-relevant features align well with steered outputs, and 15–20 features typically yield the best steering results.The paper characterizes SAE-RSV as an efficient and interpretable approach for enhancing LLM steering.Source: §6 · p. 10

ETHICAL STATEMENT

The study evaluates Llama-3-8B-Instruct and uses publicly available model-written evaluation datasets; it reports no personal-data processing or human-subjects research.

ETHICAL STATEMENT

  • The study evaluates the publicly available Llama-3-8B-Instruct model under its academic usage policy.The authors state that the model was used strictly for research.Source: ETHICAL STATEMENT · p. 10
  • The study relies on Anthropic's publicly available model-written evaluation datasets and reports compliance with dataset and model usage terms.Source: ETHICAL STATEMENT · p. 10
  • The authors report that they collected or processed no personal data, conducted no human-subjects research, and included no personally identifiable information.Source: ETHICAL STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

The reproducibility statement outlines implementation and experiment details and says code and data will be released upon acceptance. The LLM usage statement describes LLMs used as research subjects, annotators, and writing assistants.

REPRODUCIBILITY STATEMENT

  • The paper identifies pipeline, hyperparameter, and experiment details as supporting reproduction, and says code and data will be released upon acceptance.The experiment details include resources, datasets, model scale, baselines, training format, metrics, and machine-annotation prompts.Source: REPRODUCIBILITY STATEMENT · p. 10

A LLM USAGE STATEMENT

  • Llama-3-8B-Instruct serves as the research subject, with evaluation conducted under its academic usage policy.Source: A LLM USAGE STATEMENT · p. 13
  • GPT-4o-mini annotates generation quality, while ChatGPT assists with grammar, rephrasing, and visualization suggestions.The authors say they manually checked the manuscript and that it represents their original thoughts.Source: A LLM USAGE STATEMENT · p. 13

B DATASETS

The datasets provide 50 multiple-choice training pairs and 50 open-ended test questions for each of five concepts. The training format places the answer token inside instruction tags, which the authors report improves steering performance over the traditional CAA format.

B DATASETS

  • Each of five concepts has 50 multiple-choice training pairs and 50 open-ended test questions, using model-written evaluation data and GPT-4-generated datasets.The concepts are Sycophancy, Corrigibility, Hallucination, Myopic reward, and Survival instinct.Source: B DATASETS · p. 13
  • The five concepts cover agreement over factual accuracy, openness to correction, false information, immediate gains over long-term consequences, and acceptance of human deactivation or modification.Source: B DATASETS · p. 13

C FORMAT OF TRAINING DATA

  • The training format places the answer token inside instruction tags, unlike traditional CAA, and the authors report that this greatly improves steering performance.Figure 5 compares the training data formats.Source: C FORMAT OF TRAINING DATA · p. 13, Figure 5 · p. 13

D HYPERPARAMETERS

The appendix identifies task-specific SAE-RSV hyperparameter settings and describes deterministic generation and GPT-4o-mini evaluation. It also points to prompts for selecting topic-relevant features and evaluating success rates.

D HYPERPARAMETERS

E PROMPTS

  • Generation uses a 1024-token limit and greedy decoding, while GPT-4o-mini scores responses and judges whether feature semantics relate to the topic.Temperature is set to 0 to make results deterministic.Source: E PROMPTS · p. 14
  • Figure 6 shows prompts for topic-relevant feature selection, and Figure 7 shows prompts for success-rate evaluation.Source: E PROMPTS · p. 14

F ANALYSIS OF TOPIC-RELEVANT FEATURES

The feature examples align topic-relevant SAE features with target behaviors, while steered-output examples show SAE-RSV succeeding in cases where CAA fails across four concepts.

F ANALYSIS OF TOPIC-RELEVANT FEATURES

  • Examples of topic-relevant features align with target behaviors: falsehood and misinformation, feedback acceptance, agreement, and willingness or readiness for closure.Table 7 also presents examples of noising features across four concepts.Source: F ANALYSIS OF TOPIC-RELEVANT FEATURES · p. 14, Table 7 · p. 17

G EXAMPLES OF STEERED MODEL OUTPUTS

Related papers

Loading 2509.23799v2…