Papers with Sparse Autoencoders
FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Datasets Dependency (2025.acl-srw)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. |
| Approach: | They propose a method that trains SAEs on the model’s own synthetic dataset and a model-specific model to capture model-internal features. |
| Outcome: | The proposed method outperforms SAEs trained on web-based datasets and exhibits lower Fake Feature Ratio in 5 out of 7 models. |
Out of Distribution, Out of Luck: Process Rewards Misguide Reasoning Models (2026.eacl-short)
Copied to clipboard
| Challenge: | 80% of reasoning model outputs respond to formatting artifacts rather than mathematical content. |
| Approach: | They evaluate process reward models that provide step-level feedback during inference . they identify distinct reward prediction patterns that differentiate reasoning from non-reasoning model outputs . |
| Outcome: | The proposed model fails to enhance and sometimes degrade reasoning model performance. |
SFAL: Semantic-Functional Alignment Scores for Distributional Evaluation of Auto-Interpretability in Sparse Autoencoders (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Interpreting the internal representations of large language models (LLMs) is crucial for their deployment in real-world applications, impacting areas such as AI safety, debugging, and compliance. |
| Approach: | They propose an alternative evaluation strategy that assesses the alignment between the semantic neighbourhoods of features and their functional neighbourhoods by using co-occurrence statistics. |
| Outcome: | The proposed evaluation strategy reduces reliance on scoring on large-scale models and improves efficiency and cost-effectiveness. |
Evaluating the Impact of SAE-based Language Steering on LLM Performance (2026.eacl-srw)
Copied to clipboard
| Challenge: | Recent advances in Sparse Autoencoders (SAEs) have revealed interpretable features within large language models (LLMs) however, the impact of SAE-based language steering on output quality and task performance remains unclear. |
| Approach: | They apply language-specific SAE feature steering to three LLMs from two model families and evaluate it on a translation task and a multilingual question-answering task. |
| Outcome: | The proposed approach outperforms prompting and language neuron-based steering on translation and multilingual question-answering tasks. |
Probing Bias Formation in Medical LLMs through Activation Steering (2026.acl-srw)
Copied to clipboard
| Challenge: | Large Language Models specialized for the medical domain achieve high performance on static benchmarks, but are vulnerable to sycophantic confabulation. |
| Approach: | They propose a framework toward clinical AI systems that are more robust and aligned with expert medical logic. |
| Outcome: | The proposed framework outperforms static global interventions on a medical prompt with cluster-conditioned dynamic steering. |
Interpretable Company Similarity with Sparse Autoencoders (2025.acl-industry)
Copied to clipboard
Marco Molinari, Victor Shao, Luca Imeneo, Mateusz Mikolajczak, Abhimanyu Pandey, Vladimir Tregubiak, Sebastião Kuznetsov Ryder Torres Pereira
| Challenge: | Traditionally, company comparisons rely on relative returns and discrete classifications, or a combination of both. |
| Approach: | They propose to use clusters of embeddings to enhance the interpretability of Large Language Models by decomposing Large Language models activations into interpretable features. |
| Outcome: | The proposed clusters of embeddings capture the internal representation of a company description, rather than just semantic similarity alone. |
Decoding Dark Matter: Specialized Sparse Autoencoders for Interpreting Rare Concepts in Foundation Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) are a promising tool for disentangling FM representations, but they struggle to capture rare, yet crucial concepts in the data. |
| Approach: | They propose a technique to train Sparse Autoencoders to illuminate elusive dark matter features by focusing on specific subdomains. |
| Outcome: | The proposed method achieves 12.5% better classification accuracy than general-purpose SAEs when applied to remove spurious gender information. |
Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders (2025.emnlp-main)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs). |
| Approach: | They propose a method that identifies the most influential latents by incorporating output-side gradient information. |
| Outcome: | The proposed method identifies the most influential latents by incorporating output-side gradient information. |
A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) can disentangle complex features into more interpretable components. |
| Approach: | They propose to use Sparse Autoencoders to disentangle LLM features into more interpretable components. |
| Outcome: | The proposed method disentangles complex features into more interpretable components. |
Feature Drift: How Fine-Tuning Repurposes Representations in LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Sparse autoencoders (SAEs) are a powerful tool for interpreting neural networks by extracting concepts (features) represented in their activations. |
| Approach: | They propose to use Sparse Autoencoders to extract concepts from their activations to explain how fine-tuning changes model capabilities. |
| Outcome: | The proposed model recombines existing concepts rather than learning new ones, and shows that it is a better explanation for how fine-tuning changes model capabilities. |
Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse Autoencoders (2026.acl-long)
Copied to clipboard
Xiangchen Song, Aashiq Muhamed, Yujia Zheng, Lingjing Kong, Zeyu Tang, Mona T. Diab, Virginia Smith, Kun Zhang
| Challenge: | Sparse Autoencoders (SAEs) are a tool in mechanistic interpretability (MI) but the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across different training runs. |
| Approach: | They propose to use the Pairwise Dictionary Mean Correlation Coefficient to quantify SAE feature consistency as an evaluation axis alongside reconstruction and sparsity. |
| Outcome: | The proposed measure is based on the pairwise dictionary mean correlation coefficient (PW-MCC) on LLM activations. |
Unveiling Decision-Making in LLMs for Text Classification : Extraction of influential and interpretable concepts with Sparse Autoencoders (2026.findings-eacl)
Copied to clipboard
| Challenge: | Concept-based explanations for large language models are not well understood in text classification. |
| Approach: | They propose a model with a specialized classifier head and activation rate sparsity loss for sentence classification . they compare it to existing models with HI-Concept and ConceptShap . |
| Outcome: | The proposed model improves both the causality and interpretability of the extracted features. |
AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders (2026.eacl-long)
Copied to clipboard
Georgii Aparin, Tasnima Sadekova, Alexey Rukhovich, Assel Yermekova, Laida Kushnareva, Vadim Popov, Kristian Kuznetsov, Irina Piontkovskaya
| Challenge: | Feature steering reduces Whisper’s false speech detections by 70% with negligible WER increase, demonstrating real-world applicability. |
| Approach: | They train Sparse Autoencoders across all encoder layers of Whisper and HuBERT and evaluate their stability, interpretability, and practical utility. |
| Outcome: | The proposed models capture general acoustic and semantic information as well as specific events, including environmental noises and paralinguistic sounds, and disentangle them effectively. |
Sparse Latents Steer Retrieval-Augmented Generation (2025.acl-long)
Copied to clipboard
Chunlei Xin, Shuheng Zhou, Huijia Zhu, Weiqiang Wang, Xuanang Chen, Xinyan Guan, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun
| Challenge: | In this study, we uncover interpretable latents that govern RAG behavior in large language models . Sparse Autoencoders are used to control large language model (LLM) behavior . |
| Approach: | They leverage Sparse Autoencoders within the LLaMA Scope to uncover latents that govern RAG behaviors. |
| Outcome: | The proposed model can be used to control large language models without architectural modifications. |
Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) exhibit impressive abilities in various domains such as text generation, instruction following, and reasoning. |
| Approach: | They propose a method to decompose the activations of Large Language Models into a sparse linear combination of SAE features. |
| Outcome: | The proposed method shows that some features are strongly related to specific languages, while others are unaffected by ablating them. |
SAE-FiRE: Enhancing Earnings Surprise Predictions Through Sparse Autoencoder Feature Selection (2026.findings-acl)
Copied to clipboard
| Challenge: | Conference call transcripts contain significant redundancy and industry-specific terminology that creates obstacles for language models. |
| Approach: | They propose a Sparse Autoencoder for Financial Representation Enhancement framework to extract key information from earnings conference call transcripts and eliminate redundancy. |
| Outcome: | The proposed method outperforms baselines in analyzing earnings conference call transcripts. |
Deciphering Cultural Representations in Large Language Models via Sparse Autoencoders (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior work has identified so-called cultural neurons, but individual neurons are often polysemous, conflating abstract cultural knowledge with surface-level lexical cues due to superposition. |
| Approach: | They apply Sparse Autoencoders to decompose LLM activations into sparse, interpretable feature representations that disentangle culturally selective features. |
| Outcome: | The proposed model disentangles culturally selective features from paraphrasing and task formats, indicating abstraction beyond lexical correlations. |
From Insight to Action: A Novel Framework for Interpretability-Guided Data Selection in Large Language Models (2026.acl-long)
Copied to clipboard
Ling Shi, Xinwei Wu, Xiaohu Zhao, Hao Wang, Heng Liu, Yangyang Liu, Linlong Xu, Longyue Wang, Deyi Xiong, Weihua Luo
| Challenge: | Recent research in mechanistic interpretability has revealed that Large Language models contain disentangled, human-understandable components. |
| Approach: | They propose a framework that first identifies causal task features through frequency recall and interventional filtering, then selects “Feature-Resonant Data” that maximally activates task features for fine-tuning. |
| Outcome: | The proposed framework outperforms existing models on mathematical reasoning, summarization, and translation tasks while using only 50% of the data. |
Understanding the Repeat Curse in Large Language Models from a Feature Perspective (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models suffer from repetitive text generation, a phenomenon we refer to as the ”Repeat Curse”. |
| Approach: | They propose a method to induce and analyze the Repeat Curse in large language models by using mechanistic interpretability. |
| Outcome: | The proposed method induces and analyzes the Repeat Curse in large language models using mechanistic interpretability. |
Features that Make a Difference: Leveraging Gradients for Improved Dictionary Learning (2025.findings-naacl)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) are a promising approach for extracting neural network representations by learning a sparse and overcomplete decomposition of the network’s internal activations. |
| Approach: | They propose a method that learns a sparse and overcomplete decomposition of the network's internal activations and a gradient approach to learn latents. |
| Outcome: | The proposed algorithms improve the performance of the k-sparse autoencoder and the ability to learn latent features. |
SAFE: A Sparse Autoencoder-Based Framework for Robust Query Enrichment and Hallucination Mitigation in LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models suffer from hallucinations, which can undermine their performance in critical applications. |
| Approach: | They propose a framework for detecting and mitigating hallucinations by leveraging SAEs. |
| Outcome: | The proposed framework improves query generation accuracy and mitigates hallucinations across datasets. |
Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy (2025.emnlp-main)
Copied to clipboard
Nikita Balagansky, Yaroslav Aksenov, Daniil Laptev, Vadim Kurochkin, Gleb Gerasimov, Nikita Koriagin, Daniil Gavrilov
| Challenge: | Sparse Autoencoders (SAEs) are powerful tools for interpreting neural networks . conventional SAEs are constrained by the fixed sparsity level chosen during training . |
| Approach: | They propose a training objective that trains a single SAE to optimise reconstructions across multiple sparsity levels simultaneously. |
| Outcome: | The proposed objective achieves Pareto-optimal trade-offs between sparsity and explained variance, outperforming traditional SAEs trained at individual sparsities. |
SAEs Are Good for Steering – If You Select the Right Features (2025.emnlp-main)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) can learn a decomposition of a model’s latent space by analyzing the input tokens that activate them. |
| Approach: | They propose an unsupervised approach to learn a decomposition of a model’s latent space by analyzing the input tokens that activate them. |
| Outcome: | The proposed approach matches the performance of existing supervised methods by identifying features with low output scores and identifying them with input and output scores. |
Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can solve reasoning and mathematical problems using the Chain-of-Thought technique, but require costly and long CoT data and fine-tuning. |
| Approach: | They propose a method that uses Sparse Autoencoders to extract interpretable features from vanilla CoT and use them to steer the LLM's internal states. |
| Outcome: | The proposed method uses Sparse Autoencoders (SAEs) to extract interpretable features from vanilla CoT and steer the LLM's internal states during generation. |
The Mechanics of Interference: Defusing Distractors in RAG via Sparse Autoencoder Interventions (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models exhibit a critical vulnerability to distractor interference when processing retrieval-augmented contexts. |
| Approach: | They propose a mechanistic framework that corrects this failure mode through targeted interventions in the model’s latent space. |
| Outcome: | The proposed framework achieves recovery rates of up to 94% on distractor-vulnerable samples on Gemma-2 and Llama-3 model families across three QA benchmarks. |
When Truthful Representations Flip Under Deceptive Instructions? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. |
| Approach: | They use Sparse Autoencoders to analyze LLM's internal representations to determine when and how they "flip" from truthful to deceptive under deceptively crafted instructions. |
| Outcome: | The proposed model's True/False output is predictable across all conditions based on the model''s representation, and the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations. |
Uncovering Sentiment Analysis Circuit in Large Language Model (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work has shown that sentiment is encoded linearly in LLM representations, but their ability to utilize this information remains fragile to prompt variations. |
| Approach: | They propose a simple inference-time intervention method that amplifies circuit features to compensate for insufficient activation. |
| Outcome: | The proposed method improves on a sentiment analysis circuit with sparse autoencoders and circuit-level analysis. |
Controllable LLM Reasoning via Sparse Autoencoder-Based Steering (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods struggle to control fine-grained reasoning strategies due to conceptual entanglement in LRMs’ hidden states. |
| Approach: | They propose to decompose strategy-entangled hidden states into a disentangled feature space by using Sparse Autoencoders to identify the few strategy-specific features from the vast pool of SAE features. |
| Outcome: | The proposed method outperforms existing methods by 15% in control effectiveness. |
CLaS-Bench: A Cross-Lingual Alignment and Steering Benchmark (2026.findings-acl)
Copied to clipboard
Daniil Gurgurov, Yusser Al Ghussin, Tanja Baeumel, Cheng-Ting Chou, Patrick Schramowski, Marius Mosbach, Josef Van Genabith, Simon Ostermann
| Challenge: | Understanding and controlling behavior of large language models (LLMs) is an important topic in multilingual NLP. |
| Approach: | They propose a lightweight parallel-question benchmark for evaluating language-forcing behavior in large language models across 32 languages. |
| Outcome: | The proposed benchmark measures language steering in 32 languages across 32 languages. |
Towards Understanding the Robustness of Sparse Autoencoders (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are vulnerable to optimization-based jailbreak attacks that exploit internal gradient structure. |
| Approach: | They propose to integrate pretrained Sparse Autoencoders into transformer residual streams at inference time without modifying model weights or blocking gradients. |
| Outcome: | The proposed model reduces jailbreak success rate by 5x compared to baseline models . compared with models with weak white-box attacks, the proposed model is more robust . |
Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders (2025.findings-acl)
Copied to clipboard
Kristian Kuznetsov, Laida Kushnareva, Anton Razzhigaev, Polina Druzhinina, Anastasia Voznyuk, Irina Piontkovskaya, Evgeny Burnaev, Serguei Barannikov
| Challenge: | Existing algorithms for AI text detection lack interpretability, limiting their reliability in highstakes applications. |
| Approach: | They extend existing ATD frameworks by using Sparse Autoencoders to extract features from Gemma-2-2b residual stream. |
| Outcome: | The proposed algorithms can extract human-interpretable features from Gemma-2-2b model. |
Sparse Autoencoder Features for Classifications and Transferability (2025.emnlp-main)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) provide potential for uncovering structured, human-interpretable representations in Large Language Models (LLMs). |
| Approach: | They analyze SAEs for interpretable feature extraction from Large Language Models in safety-critical classification tasks. |
| Outcome: | The proposed framework outperforms hidden-state and BoW models while demonstrating cross-lingual toxicity detection and visual classification tasks. |
Multilingual Language Models Encode Script Over Linguistic Structure (2026.acl-long)
Copied to clipboard
| Challenge: | a recent study suggests that multilingual language models organize representations around surface form, but the nature of this internal organization remains elusive. |
| Approach: | They analyze language-associated units across different model families and scales . romanization induces near-disjoint representations that align with neither native-script inputs nor English . |
| Outcome: | The results show that multilingual language models organize representations around surface form . romanization induces near-disjoint representations that align with neither native-script inputs nor English . |
Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models (2026.acl-long)
Copied to clipboard
Woody Haosheng Gan, Deqing Fu, Julian Asilis, Ollie Liu, Vatsal Sharan, Robin Jia, Willie Neiswanger
| Challenge: | Steering methods have emerged as effective tools for guiding large language models’ behavior, yet multimodal large language model lacks comparable techniques due to architectural diversity and limited availability of multimodal steering vectors. |
| Approach: | They validate steering vectors derived solely from text-only LLM backbones and use a cross-modal transfer technique to reuse existing interpretability tools. |
| Outcome: | The proposed steering vectors can guide and enhance multimodal models using SPAR, Mean Shift, and Linear Probing. |