A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders

• URL: http://arxiv.org/abs/2409.14507v5
• Date: Mon, 02 Jun 2025 10:58:16 GMT
• Title: A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders
• Authors: David Chanin, James Wilken-Smith, Tomáš Dulka, Hardik Bhatnagar, Satvik Golechha, Joseph Bloom,
• Abstract summary: We show that sparse decomposition and splitting of hierarchical features is not robust.<n>Specifically, we show that seemingly monosemantic features fail to fire where they should, and instead get "absorbed" into their children features.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Sparse Autoencoders (SAEs) aim to decompose the activation space of large language models (LLMs) into human-interpretable latent directions or features. As we increase the number of features in the SAE, hierarchical features tend to split into finer features ("math" may split into "algebra", "geometry", etc.), a phenomenon referred to as feature splitting. However, we show that sparse decomposition and splitting of hierarchical features is not robust. Specifically, we show that seemingly monosemantic features fail to fire where they should, and instead get "absorbed" into their children features. We coin this phenomenon feature absorption, and show that it is caused by optimizing for sparsity in SAEs whenever the underlying features form a hierarchy. We introduce a metric to detect absorption in SAEs, and validate our findings empirically on hundreds of LLM SAEs. Our investigation suggests that varying SAE sizes or sparsity is insufficient to solve this issue. We discuss the implications of feature absorption in SAEs and some potential approaches to solve the fundamental theoretical issues before SAEs can be used for interpreting LLMs robustly and at scale.

Related papers

• Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs [100.02824137397464]
  We investigate how Large Language Models adapt their internal representations when encountering inputs of increasing difficulty.<n>We reveal a consistent and quantifiable phenomenon: as task difficulty increases, the last hidden states of LLMs become substantially sparser.<n>This sparsity--difficulty relation is observable across diverse models and domains.
  arXiv  Detail & Related papers  (2026-03-03T18:48:15Z)
• Step-Level Sparse Autoencoder for Reasoning Process Interpretation [48.99201531966593]
  Large Language Models (LLMs) have achieved strong complex reasoning capabilities through Chain-of-Thought (CoT) reasoning.<n>We propose step-level sparse autoencoder (SSAE), which serves as an analytical tool to disentangle different aspects of LLMs' reasoning steps into sparse features.<n> Experiments on multiple base models and reasoning tasks show the effectiveness of the extracted features.
  arXiv  Detail & Related papers  (2026-03-03T14:25:02Z)
• Sparse Semantic Dimension as a Generalization Certificate for LLMs [53.681678236115836]
  We introduce the Sparse Semantic Dimension (SSD), a complexity measure derived from the active feature vocabulary of a Sparse Autoencoder (SAE) trained on the model's layers.<n>We validate this framework on GPT-2 Small and Gemma-2B, demonstrating that our bound provides non-vacuous certificates at realistic sample sizes.
  arXiv  Detail & Related papers  (2026-02-11T21:45:18Z)
• Beyond Redundancy: Diverse and Specialized Multi-Expert Sparse Autoencoder [59.89996751196727]
  Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting large language models.<n>SAEs' hidden layers have high dimensionality to satisfy sparsity constraints, resulting in prohibitive training and inference costs.<n>Recent Mixture of Experts (MoE) approaches attempt to address this by SAEs into narrower expert networks with gated activation.<n>We propose two key innovations: (1) Multiple Expert Activation that simultaneously engages semantically weighted expert subsets to encourage specialization, and (2) Feature Scaling that enhances diversity through adaptive high-frequency scaling.
  arXiv  Detail & Related papers  (2025-11-07T22:19:34Z)
• ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders [30.219733023958188]
  Sparse Autoencoder (SAE) has emerged as a powerful tool for mechanistic interpretability of large language models.<n>We propose a semantically-guided SAE, called ProtSAE.<n>We show that ProtSAE learns more biologically relevant and interpretable hidden features compared to previous methods.
  arXiv  Detail & Related papers  (2025-08-26T11:20:31Z)
• LIRA: Inferring Segmentation in Large Multi-modal Models with Local Interleaved Region Assistance [56.474856189865946]
  Large multi-modal models (LMMs) struggle with inaccurate segmentation and hallucinated comprehension.<n>We propose LIRA, a framework that capitalizes on the complementary relationship between visual comprehension and segmentation.<n>LIRA achieves state-of-the-art performance in both segmentation and comprehension tasks.
  arXiv  Detail & Related papers  (2025-07-08T07:46:26Z)
• Dense SAE Latents Are Features, Not Bugs [75.08462524662072]
  We show that dense latents serve functional roles in language model computation.<n>We identify classes tied to position tracking, context binding, entropy regulation, letter-specific output signals, part-of-speech, and principal component reconstruction.
  arXiv  Detail & Related papers  (2025-06-18T17:59:35Z)
• Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders [1.0582505915332336]
  It is assumed that sparse autoencoders (SAEs) decompose polysemantic activations into interpretable linear directions.<n>We find that if an SAE is more narrow than the number of underlying "true features" on which it is trained, and there is correlation between features, the SAE will merge components of correlated features together.<n>This phenomenon, which we call feature hedging, is caused by SAE reconstruction loss, and is more severe the narrower the SAE.
  arXiv  Detail & Related papers  (2025-05-16T23:30:17Z)
• Empirical Evaluation of Progressive Coding for Sparse Autoencoders [45.94517951918044]
  We show that dictionary importance in vanilla SAEs follows a power law.<n>We show Matryoshka SAEs exhibit lower reconstruction loss and recaptured language modeling loss.
  arXiv  Detail & Related papers  (2025-04-30T21:08:32Z)
• Evaluating and Designing Sparse Autoencoders by Approximating Quasi-Orthogonality [3.9230690073443166]
  We show that the magnitude of sparse feature vectors can be approximated using their corresponding dense vector with a closed-form error bound.<n>We introduce Approximate Activation Feature (AFA), which approximates the magnitude of the ground-truth sparse feature vector.<n>We demonstrate that top-AFA SAEs achieve reconstruction loss comparable to that of state-of-the-art top-k SAEs.
  arXiv  Detail & Related papers  (2025-03-31T16:22:11Z)
• Do Sparse Autoencoders Generalize? A Case Study of Answerability [12.131254862319865]
  We evaluate SAE feature generalization across diverse answerability datasets for Gemma 2 SAEs. Our analysis reveals that residual stream probes outperform SAE features within domains, but generalization performance differs sharply.
  arXiv  Detail & Related papers  (2025-02-27T10:45:25Z)
• Sparse Autoencoders Do Not Find Canonical Units of Analysis [6.0188420022822955]
  A common goal of mechanistic interpretability is to decompose the activations of neural networks into features. Sparse autoencoders (SAEs) are a popular method for finding these features. We use two techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic.
  arXiv  Detail & Related papers  (2025-02-07T12:33:08Z)
• Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words [29.09237503747052]
  Sparse autoencoders (SAEs) have gained a lot of attention as a promising tool to improve the interpretability of large language models (LLMs) In this paper, we propose a suite of evaluations for SAEs to analyze the quality of monosemantic features by focusing on polysemous words.
  arXiv  Detail & Related papers  (2025-01-09T02:54:19Z)
• Content-Style Learning from Unaligned Domains: Identifiability under Unknown Latent Dimensions [11.872336932802844]
  We introduce a new analytical framework via cross-domain textitlatent distribution matching (LDM) We show that restrictive assumptions such as component-wise independence of the latent variables can be removed. We recast the LDM formulation into a regularized multi-domain GAN loss with coupled latent variables.
  arXiv  Detail & Related papers  (2024-11-06T08:30:23Z)
• Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders [115.34050914216665]
  Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models. We introduce a suite of 256 SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. We assess the generalizability of SAEs trained on base models to longer contexts and fine-tuned models.
  arXiv  Detail & Related papers  (2024-10-27T17:33:49Z)
• Automatically Interpreting Millions of Features in Large Language Models [1.8035046415192353]
  sparse autoencoders (SAEs) can be used to transform activations into a higher-dimensional latent space.<n>We build an open-source pipeline to generate and evaluate natural language explanations for SAE features.<n>Our large-scale analysis confirms that SAE latents are indeed much more interpretable than neurons.
  arXiv  Detail & Related papers  (2024-10-17T17:56:01Z)
• Interpretability as Compression: Reconsidering SAE Explanations of Neural Activations with MDL-SAEs [0.0]
  We present an information-theoretic framework for interpreting SAEs as lossy compression algorithms. We argue that using MDL rather than sparsity may avoid potential pitfalls with naively maximising sparsity.
  arXiv  Detail & Related papers  (2024-10-15T01:38:03Z)
• Disentangling Dense Embeddings with Sparse Autoencoders [0.0]
  Sparse autoencoders (SAEs) have shown promise in extracting interpretable features from complex neural networks. We present one of the first applications of SAEs to dense text embeddings from large language models. We show that the resulting sparse representations maintain semantic fidelity while offering interpretability.
  arXiv  Detail & Related papers  (2024-08-01T15:46:22Z)
• Interpreting Attention Layer Outputs with Sparse Autoencoders [3.201633659481912]
  Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We show that Sparse Autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work.
  arXiv  Detail & Related papers  (2024-06-25T17:43:13Z)
• Crafting Interpretable Embeddings by Asking LLMs Questions [89.49960984640363]
  Large language models (LLMs) have rapidly improved text embeddings for a growing array of natural-language processing tasks. We introduce question-answering embeddings (QA-Emb), embeddings where each feature represents an answer to a yes/no question asked to an LLM. We use QA-Emb to flexibly generate interpretable models for predicting fMRI voxel responses to language stimuli.
  arXiv  Detail & Related papers  (2024-05-26T22:30:29Z)
• Aligning Language Models to Explicitly Handle Ambiguity [22.078095273053506]
  We propose Alignment with Perceived Ambiguity (APA), a novel pipeline that aligns language models to deal with ambiguous queries. Experimental results on question-answering datasets demonstrate that APA empowers LLMs to explicitly detect and manage ambiguous queries. Our finding proves that APA excels beyond training with gold-standard labels, especially in out-of-distribution scenarios.
  arXiv  Detail & Related papers  (2024-04-18T07:59:53Z)
• SEMQA: Semi-Extractive Multi-Source Question Answering [94.04430035121136]
  We introduce a new QA task for answering multi-answer questions by summarizing multiple diverse sources in a semi-extractive fashion. We create the first dataset of this kind, QuoteSum, with human-written semi-extractive answers to natural and generated questions.
  arXiv  Detail & Related papers  (2023-11-08T18:46:32Z)
• Interpretability at Scale: Identifying Causal Mechanisms in Alpaca [62.65877150123775]
  We use Boundless DAS to efficiently search for interpretable causal structure in large language models while they follow instructions. Our findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models.
  arXiv  Detail & Related papers  (2023-05-15T17:15:40Z)
• We're Afraid Language Models Aren't Modeling Ambiguity [136.8068419824318]
  Managing ambiguity is a key part of human language understanding. We characterize ambiguity in a sentence by its effect on entailment relations with another sentence. We show that a multilabel NLI model can flag political claims in the wild that are misleading due to ambiguity.
  arXiv  Detail & Related papers  (2023-04-27T17:57:58Z)
• A Latent-Variable Model for Intrinsic Probing [93.62808331764072]
  We propose a novel latent-variable formulation for constructing intrinsic probes. We find empirical evidence that pre-trained representations develop a cross-lingually entangled notion of morphosyntax.
  arXiv  Detail & Related papers  (2022-01-20T15:01:12Z)
• Self-Supervised Learning Disentangled Group Representation as Feature [82.07737719232972]
  We show that existing Self-Supervised Learning (SSL) only disentangles simple augmentation features such as rotation and colorization. We propose an iterative SSL algorithm: Iterative Partition-based Invariant Risk Minimization (IP-IRM) We prove that IP-IRM converges to a fully disentangled representation and show its effectiveness on various benchmarks.
  arXiv  Detail & Related papers  (2021-10-28T16:12:33Z)
• Where and What? Examining Interpretable Disentangled Representations [96.32813624341833]
  Capturing interpretable variations has long been one of the goals in disentanglement learning. Unlike the independence assumption, interpretability has rarely been exploited to encourage disentanglement in the unsupervised setting. In this paper, we examine the interpretability of disentangled representations by investigating two questions: where to be interpreted and what to be interpreted.
  arXiv  Detail & Related papers  (2021-04-07T11:22:02Z)
• Dive into Ambiguity: Latent Distribution Mining and Pairwise Uncertainty Estimation for Facial Expression Recognition [59.52434325897716]
  We propose a solution, named DMUE, to address the problem of annotation ambiguity from two perspectives. For the former, an auxiliary multi-branch learning framework is introduced to better mine and describe the latent distribution in the label space. For the latter, the pairwise relationship of semantic feature between instances are fully exploited to estimate the ambiguity extent in the instance space.
  arXiv  Detail & Related papers  (2021-04-01T03:21:57Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.