Sparse autoencoders (SAEs) are emerging as a key analytical tool in the field of mechanistic interpretability for large language models (LLMs). While SAEs have primarily been used for interpretability, we shift focus and explore an understudied question: “Can SAEs be applied to practical tasks beyond interpretability?” Given that SAEs are trained on billions of tokens for sparse reconstruction, we believe they can serve as effective extractors, offering a wide range of useful knowledge that can benefit practical applications. Building on this motivation, we demonstrate that SAEs can be effectively applied to in-context learning (ICL). In particular, we highlight the utility of the SAE-reconstruction loss by showing that it provides a valuable signal in ICL—exhibiting a strong correlation with LLM performance and offering a powerful unsupervised approach for prompt selection. These findings underscore the versatility of SAEs and reveal their potential for real-world applications beyond interpretability. Our code is available at https://github.com/ihcho2/SAE-GPS.
论文检索
输入标题、作者或关键词，从 3,199 篇学术成果中精准定位
Instruction fine-tuning enhances the alignment of autoregressive language models (ArLMs) with human intent but relies on large-scale annotated datasets prone to label and text noise. In this paper, we show that existing noise detection techniques designed for autoencoder models (AeLMs) do not directly generalize to ArLMs due to differences in learning dynamics. We propose TDRanker, a novel approach leveraging training dynamics to rank datapoints from easy-to-learn to hard-to-learn, effectively identifying noisy instances. Our method demonstrates robustness across multiple model architectures covering both autoencoder and autoregressive language models (GPT-2, BERT, LaMini-Cerebras-256M) and across various dataset noise levels, achieving at least 2x faster denoising than previous techniques. Applied to real-world classification and generative tasks, TDRanker significantly improves data quality and model performance. These findings suggest that TDRanker provides a scalable solution for refining instruction-tuning datasets, enhancing the reliability of fine-tuned ArLMs in practical applications.
How a Bilingual LM Becomes Bilingual: Tracing Internal Representations with Sparse Autoencoders
PDF ↗This study explores how bilingual language models develop complex internal representations.We employ sparse autoencoders to analyze internal representations of bilingual language models with a focus on the effects of training steps, layers, and model sizes.Our analysis shows that language models first learn languages separately, and then gradually form bilingual alignments, particularly in the mid layers. We also found that this bilingual tendency is stronger in larger models.Building on these findings, we demonstrate the critical role of bilingual representations in model performance by employing a novel method that integrates decomposed representations from a fully trained model into a mid-training model.Our results provide insights into how language models acquire bilingual capabilities.
Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks. However, they still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications. Existing approaches to address this issue focus on incorporating external knowledge bases, alignment training, or decoding strategies, all of which require substantial computational cost and time. Recent works try to explore more efficient alternatives by adjusting LVLMs’ internal representations. Although promising, these methods may cause hallucinations to be insufficiently suppressed or lead to excessive interventions that negatively affect normal semantics. In this work, we leverage sparse autoencoders (SAEs) to identify semantic directions closely associated with faithfulness or hallucination, extracting more precise and disentangled hallucination-related representations. Our analysis demonstrates that interventions along the identified faithful direction can mitigate hallucinations, while those along the hallucinatory direction can exacerbate them. Building on these insights, we propose **S**teering LVLMs via **S**AE **L**atent Directions (SSL), a plug-and-play method based on SAE-derived latent directions to mitigate hallucinations in LVLMs. Extensive experiments demonstrate that SSL significantly outperforms existing decoding approaches in mitigating hallucinations, while maintaining transferability across different model architectures with negligible additional time overhead. The code is available at [https://github.com/huazhenglin2003/SSL](https://github.com/huazhenglin2003/SSL).
SAFE: A Sparse Autoencoder-Based Framework for Robust Query Enrichment and Hallucination Mitigation in LLMs
PDF ↗Despite the state-of-the-art performance of Large Language Models (LLMs), these models often suffer from hallucinations, which can undermine their performance in critical applications. In this work, we propose SAFE, a novel framework for detecting and mitigating hallucinations by leveraging Sparse Autoencoders (SAEs). While hallucination detection techniques and SAEs have been explored independently, their synergistic application in a comprehensive system, particularly for hallucination-aware query enrichment, has not been fully investigated. To validate the effectiveness of SAFE, we evaluate it on two models with available SAEs across four diverse cross-domain datasets designed to assess hallucination problems. Empirical results demonstrate that SAFE consistently improves query generation accuracy and mitigates hallucinations across all datasets, achieving accuracy improvements of up to 29.45%.
Refusal is a key safety behavior in aligned language models, yet the internal mechanisms driving refusals remain opaque. In this work, we conduct a mechanistic study of refusal in instruction-tuned LLMs using sparse autoencoders to identify latent features that causally mediate refusal behaviors. We apply our method to two open-source chat models and intervene on refusal-related features to assess their influence on generation, validating their behavioral impact across multiple harmful datasets. This enables a fine-grained inspection of how refusal manifests at the activation level and addresses key research questions such as investigating upstream-downstream latent relationship and understanding the mechanisms of adversarial jailbreaking techniques. We also establish the usefulness of refusal features in enhancing generalization for linear probes to out-of-distribution adversarial samples in classification tasks.
A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models
PDF ↗Large Language Models (LLMs) have transformed natural language processing, yet their internal mechanisms remain largely opaque. Recently, mechanistic interpretability has attracted significant attention from the research community as a means to understand the inner workings of LLMs. Among various mechanistic interpretability approaches, Sparse Autoencoders (SAEs) have emerged as a promising method due to their ability to disentangle the complex, superimposed features within LLMs into more interpretable components. This paper presents a comprehensive survey of SAEs for interpreting and understanding the internal workings of LLMs. Our major contributions include: (1) exploring the technical framework of SAEs, covering basic architecture, design improvements, and effective training strategies; (2) examining different approaches to explaining SAE features, categorized into input-based and output-based explanation methods; (3) discussing evaluation methods for assessing SAE performance, covering both structural and functional metrics; and (4) investigating real-world applications of SAEs in understanding and manipulating LLM behaviors.
Role-playing has emerged as an effective technique for enhancing the reasoning capabilities of large language models (LLMs). However, existing methods primarily rely on prompt engineering, which often lacks stability and interpretability. In this paper, we introduce Sparse Autoencoder Role-Playing Steering (SRPS), a novel framework that identifies and manipulates internal model features associated with role-playing behavior. Our approach extracts latent representations from role-play prompts, selects the most relevant features based on activation patterns, and constructs a steering vector that can be injected into the model’s residual stream with controllable intensity. Our method enables fine-grained control over role-specific behavior and offers insights into how role information influences internal model activations. Extensive experiments across various reasoning benchmarks and model sizes demonstrate consistent performance gains. Notably, in the zero-shot chain-of-thought (CoT) setting, the accuracy of Llama3.1-8B on CSQA improves from 31.86% to 39.80%, while Gemma2-9B on SVAMP increases from 37.50% to 45.10%. These results highlight the potential of SRPS to enhance reasoning ability in LLMs, providing better interpretability and stability compared to traditional prompt-based role-playing.
Large language models (LLMs) require continual knowledge updates to keep pace with the evolving world. While various model editing methods have been proposed, most face critical challenges in the context of lifelong learning due to two fundamental limitations: (1) Edit Overshooting - parameter updates intended for a specific fact spill over to unrelated regions, causing interference with previously retained knowledge; and (2) Knowledge Entanglement - polysemantic neurons’ overlapping encoding of multiple concepts makes it difficult to isolate and edit a single fact. In this paper, we propose MicroEdit, a neuron-level editing method that performs minimal and controlled interventions within LLMs. By leveraging a sparse autoencoder (SAE), MicroEdit disentangles knowledge representations and activates only a minimal set of necessary neurons for precise parameter updates. This targeted design enables fine-grained control over the editing scope, effectively mitigating interference and preserving unrelated knowledge. Extensive experiments show that MicroEdit outperforms prior methods and robustly handles lifelong knowledge editing across QA and Hallucination settings on LLaM and Mistral.
Sparse Autoencoders (SAEs) provide potential for uncovering structured, human-interpretable representations in Large Language Models (LLMs), making them a crucial tool for transparent and controllable AI systems. We systematically analyze SAE for interpretable feature extraction from LLMs in safety-critical classification tasks. Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations. SAE-derived features achieve macro F1 > 0.8, outperforming hidden-state and BoW baselines while demonstrating cross-model transfer from Gemma 2 2B to 9B-IT models. These features generalize in a zero-shot manner to cross-lingual toxicity detection and visual classification tasks. Our analysis highlights the significant impact of pooling strategies and binarization thresholds, showing that binarization offers an efficient alternative to traditional feature selection while maintaining or improving performance. These findings establish new best practices for SAE-based interpretability and enable scalable, transparent deployment of LLMs in real-world applications.
Mechanistic interpretability research faces a gap between analyzing simple circuits in toy tasks and discovering features in large models. To bridge this gap, we propose text-to-SQL generation as an ideal task to study, as it combines the formal structure of toy tasks with real-world complexity. We introduce TinySQL, a synthetic dataset, progressing from basic to advanced SQL operations, and train models ranging from 33M to 1B parameters to establish a comprehensive testbed for interpretability. We apply multiple complementary interpretability techniques, including Edge Attribution Patching and Sparse Autoencoders, to identify minimal circuits and components supporting SQL generation. We compare circuits for different SQL subskills, evaluating their minimality, reliability, and identifiability. Finally, we conduct a layerwise logit lens analysis to reveal how models compose SQL queries across layers: from intent recognition to schema resolution to structured generation. Our work provides a robust framework for probing and comparing interpretability methods in a structured, progressively complex setting.
Toward Efficient Sparse Autoencoder-Guided Steering for Improved In-Context Learning in Large Language Models
PDF ↗Sparse autoencoders (SAEs) have emerged as a powerful analytical tool in mechanistic interpretability for large language models (LLMs), with growing success in applications beyond interpretability. Building on this momentum, we present a novel approach that leverages SAEs to enhance the general in-context learning (ICL) performance of LLMs.Specifically, we introduce Feature Detection through Prompt Variation (FDPV), which leverages the SAE’s remarkable ability to capture subtle differences between prompts, enabling efficient feature selection for downstream steering. In addition, we propose a novel steering method tailored to ICL—Selective In-Context Steering (SISTER)—grounded in recent insights from ICL research that LLMs utilize label words as key anchors. Our method yields a 3.5% average performance improvement across diverse text classification tasks and exhibits greater robustness to hyperparameter variations compared to standard steering approaches. Our code is available at https://github.com/ihcho2/SAE-ICL.
Large language models (LLMs) store vast amounts of information, making them powerful yet raising privacy and safety concerns when selective knowledge removal is required. Existing unlearning strategies, ranging from gradient-based fine-tuning and model editing to sparse autoencoder (SAE) steering, either lack interpretability or fail to provide a robust defense against adversarial prompts. We propose **S**AE–Guided **S**ubspace **P**rojection **U**nlearning (**SSPU**), a novel framework that leverages SAE features to drive targeted updates in the model’s parameter space, enabling precise, interpretable, and robust unlearning. SSPU’s three-stage pipeline performs data-driven layer and feature selection, subspace construction via QR decomposition, and constrained optimization that controls activations into an “irrelevant” subspace while preserving retained knowledge. Overall, we use SAE features to construct a subspace that supervises unlearning, refining the loss and adding a regularization term to guide interpretable parameter updates. In experiments on the WMDP–Cyber forget set and three utility benchmarks (MMLU, TruthfulQA, GSM8K), SSPU reduces harmful knowledge accuracy by 3.22% compared to the strongest baseline. It also improves adversarial robustness, lowering malicious accuracy under jailbreak prompts compared to baselines. Our findings expose the limitations of prior unlearning methods and demonstrate how interpretable subspace-guided optimization can achieve robust, controllable model behavior.
Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval
PDF ↗Despite their strong performance, Dense Passage Retrieval (DPR) models suffer from a lackof interpretability. In this work, we propose a novel interpretability framework that leveragesSparse Autoencoders (SAEs) to decompose previously uninterpretable dense embeddings fromDPR models into distinct, interpretable latent concepts. We generate natural language descriptionsfor each latent concept, enabling human interpretations of both the dense embeddingsand the query-document similarity scores of DPR models. We further introduce Concept-Level Sparse Retrieval (CL-SR), a retrieval framework that directly utilizes the extractedlatent concepts as indexing units. CL-SR effectively combines the semantic expressiveness ofdense embeddings with the transparency and efficiency of sparse representations. We showthat CL-SR achieves high index-space and computational efficiency while maintaining robustperformance across vocabulary and semantic mismatches.
Debiasing techniques such as SentDebias aim to reduce bias in large language models (LLMs). Previous studies have evaluated their cross-lingual transferability by directly applying these methods to LLM representations, revealing their limited effectiveness across languages. In this work, we therefore propose to perform debiasing in a joint latent space rather than directly on LLM representations. We construct a well-aligned cross-lingual latent space using an autoencoder trained on parallel TED talk scripts. Our experiments with Aya-expanse and two debiasing techniques across four languages (English, French, German, Dutch) demonstrate that a) autoencoders effectively construct a well-aligned cross-lingual latent space, and b) applying debiasing techniques in the learned cross-lingual latent space significantly improves both the overall debiasing performance and cross-lingual transferability.
Group-SAE: Efficient Training of Sparse Autoencoders for Large Language Models via Layer Groups
PDF ↗Sparse AutoEncoders (SAEs) have recently been employed as a promising unsupervised approach for understanding the representations of layers of Large Language Models (LLMs). However, with the growth in model size and complexity, training SAEs is computationally intensive, as typically one SAE is trained for each model layer. To address such limitation, we propose Group-SAE, a novel strategy to train SAEs. Our method considers the similarity of the residual stream representations between contiguous layers to group similar layers and train a single SAE per group. To balance the trade-off between efficiency and performance, we further introduce AMAD (Average Maximum Angular Distance), an empirical metric that guides the selection of an optimal number of groups based on representational similarity across layers. Experiments on models from the Pythia family show that our approach significantly accelerates training with minimal impact on reconstruction quality and comparable downstream task performance and interpretability over baseline SAEs trained layer by layer. This method provides an efficient and scalable strategy for training SAEs in modern LLMs.
Concept editing aims to control specific concepts in large language models (LLMs) and is an emerging subfield of model editing. Despite the emergence of various editing methods in recent years, there remains a lack of rigorous theoretical analysis and a unified perspective to systematically understand and compare these methods. To address this gap, we propose a unified paradigm for concept editing methods, in which all forms of conceptual injection are aligned at the neuron level. We study four representative concept editing methods: Neuron Editing (NE), Supervised Fine-tuning (SFT), Sparse Autoencoder (SAE), and Steering Vector (SV). Then we categorize them into two classes based on their mode of conceptual information injection: indirect (NE, SFT) and direct (SAE, SV). We evaluate above methods along four dimensions: editing reliability, output generalization, neuron level consistency, and mathematical formalization. Experiments show that SAE achieves the best editing reliability. In output generalization, SAE captures features closer to human-understood concepts, while NE tends to locate text patterns rather than true semantics. Neuron-level analysis reveals that direct methods share high neuron overlap, as do indirect methods, indicating methodological commonality within each category. Our unified paradigm offers a clear framework and valuable insights for advancing interpretability and controlled generation in LLMs.
Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis. To bridge this gap, we investigate when and how these representations “flip”, such as from truthful to deceptive, under deceptive versus truthful/neutral instructions. Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model’s instructed True/False output is predictable via linear probes across all conditions based on the internal representation. Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets. We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces.
Topic modeling is a powerful unsupervised tool for knowledge discovery. However, existing work struggles with generating limited-quality topics that are uninformative and incoherent, which hindering interpretable insights from managing textual data. In this paper, we improve the original variational autoencoder framework by incorporating contextual and graph information to address the above issues. First, the encoder utilizes topic fusion techniques to combine contextual and bag-of-words information well, and meanwhile exploits the constraints of topic alignment and topic sharpening to generate informative topics. Second, we develop a simple word co-occurrence graph information fusion strategy that efficiently increases topic coherence. On three benchmark datasets, our new framework generates more coherent and diverse topics compared to various baselines, and achieves strong performance on both automatic and manual evaluations.
Large language models (LLMs) are now ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs, including profanity, vulgarity, and derogatory remarks. Although numerous detoxification methods exist, most apply broad, surface-level fixes and can therefore easily be circumvented by jailbreak attacks. In this paper we leverage sparse autoencoders (SAEs) to identify toxicity-related directions in the residual stream of models and perform targeted activation steering using the corresponding decoder vectors. We introduce three tiers of steering aggressiveness and evaluate them on GPT-2 Small and Gemma-2-2B, revealing trade-offs between toxicity reduction and language fluency. At stronger steering strengths, these causal interventions surpass competitive baselines in reducing toxicity by up to 20%, though fluency can degrade noticeably on GPT-2 Small depending on the aggressiveness. Crucially, standard NLP benchmark scores upon steering remain stable, indicating that the model’s knowledge and general abilities are preserved. We further show that feature-splitting in wider SAEs hampers safety interventions, underscoring the importance of disentangled feature learning. Our findings highlight both the promise and the current limitations of SAE-based causal interventions for LLM detoxification, further suggesting practical guidelines for safer language-model deployment.