mlsanityData to 7 Oct 2026

Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words

arXiv 2501.06254 ↗First posted 9 Jan 2025cs.CL cs.AI cs.LGcited by 17

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

Sparse autoencoders (SAEs) have gained a lot of attention as a promising tool to improve the interpretability of large language models (LLMs) by mapping the complex superposition of polysemantic neurons into monosemantic features and composing a sparse dictionary of words. However, traditional performance metrics like Mean Squared Error and L0 sparsity ignore the evaluation of the semantic representational power of SAEs -- whether they can acquire interpretable monosemantic features while preserving the semantic relationship of words. For instance, it is not obvious whether a learned sparse feature could distinguish different meanings in one word. In this paper, we propose a suite of evaluations for SAEs to analyze the quality of monosemantic features by focusing on polysemous words. Our findings reveal that SAEs developed to improve the MSE-L0 Pareto frontier may confuse interpretability, which does not necessarily enhance the extraction of monosemantic features. The analysis of SAEs with polysemous words can also figure out the internal mechanism of LLMs; deeper layers and the Attention module contribute to distinguishing polysemy in a word. Our semantics focused evaluation offers new insights into the polysemy and the existing SAE objective and contributes to the development of more practical SAEs.

Key facts. Cited by 17 papers in the map (data through 7 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 26 of its 37 references that are in the map. First posted on arXiv 9 Jan 2025.

Builds on

26 of 37 references.

References · 26

  1. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,734
  2. 2019SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systemscited by 1,718
  3. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
  4. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,130
  5. 2023Progress measures for grokking via mechanistic interpretabilitycited by 822
  6. 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetscited by 599
  7. 2024Scaling and evaluating sparse autoencoderscited by 530
  8. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 415
  9. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 381
  10. 2018WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representationscited by 370
  11. 2023Finding Neurons in a Haystack: Case Studies with Sparse Probingcited by 328
  12. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 283
14 more
  1. 2024RedPajama: an Open Dataset for Training Large Language Modelscited by 178
  2. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 171
  3. 2023Does Circuit Analysis Interpretability Scale? Evidence from Multiple Choice Capabilities in Chinchillacited by 145
  4. 2024Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Controlcited by 83
  5. 2024The Geometry of Concepts: Sparse Autoencoder Feature Structurecited by 52
  6. 2024Interpreting Attention Layer Outputs with Sparse Autoencoderscited by 43
  7. 2020XL-WiC: A Multilingual Benchmark for Evaluating Semantic Contextualizationcited by 42
  8. 2024Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Smallcited by 34
  9. 2024One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Modelscited by 30
  10. 2024IllusionVQA: A Challenging Optical Illusion Dataset for Vision Language Modelscited by 29
  11. 2024Dictionary Learning Improves Patch-Free Circuit Discovery in Mechanistic Interpretability: A Case Study on Othello-GPTcited by 27
  12. 2019JumpReLU: A Retrofit Defense Strategy for Adversarial Attackscited by 14
  13. 2024Towards Empirical Interpretation of Internal Circuits and Properties in Grokked Transformers on Modular Polynomialscited by 14
  14. 2024Evaluating Sparse Autoencoders on Targeted Concept Erasure Taskscited by 14

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 8; 2026: 9.

17 papers.

Sort:

By area:

  1. 2025On the Limits of Sparse Autoencoders: A Theoretical Framework and Reweighted Remedycited by 12
  2. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?cited by 10
  3. 2025Probing the Representational Power of Sparse Autoencoders in Vision Modelscited by 9
  4. 2026The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Levelcited by 7
  5. 2025Evaluating and Designing Sparse Autoencoders by Approximating Quasi-Orthogonalitycited by 6
  6. 2025Hypothesis-Driven Feature Manifold Analysis in LLMs via Supervised Multi-Dimensional Scalingcited by 6
  7. 2026Understanding Emergent Misalignment via Feature Superposition Geometrycited by 6
  8. 2025AlignSAE: Concept-Aligned Sparse Autoencoderscited by 4
  9. 2025Evaluating Sparse Autoencoders for Monosemantic Representationcited by 3
  10. 2025Distributional Semantics Tracing: A Framework for Explaining Hallucinations in Large Language Modelscited by 3
  11. 2026The Rate-Distortion-Polysemanticity Tradeoff in SAEscited by 1
  12. 2025Concept-SAE: A Controllable and Invertible Concept Interface for Sparse Autoencoderscited by 0
5 more
  1. 2026Unveiling the Reasoning Process of Large Language Modelscited by 0
  2. 2026Polysemanticity or Polysemy? Lexical Identity Confounds Superposition Metricscited by 0
  3. 2026Building Fast, Evaluating Slow: Pipeline Choices Dominate Autointerpretability Score Variancecited by 0
  4. 2026Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effectscited by 0
  5. 2026Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Trainingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features
  2. 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models
  3. 2025Exploring Task Performance with Interpretable Models via Sparse Auto-Encoders
  4. 2026Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherence
  5. 2026Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models
  6. 2025Beyond Redundancy: Diverse and Specialized Multi-Expert Sparse Autoencoder
  7. 2026From Input to Output: A Flexible Agent for Dual-End Interpretation of Sparse Autoencoder Features
  8. 2025Interpreting the linear structure of vision-language model embedding spaces