mlsanityData to 7 Oct 2026

Do Sparse Autoencoders Generalize? A Case Study of Answerability

arXiv 2502.19964 ↗First posted 27 Feb 2025cs.LGcited by 8

Not on the map. Why?

Abstract (arXiv)

Sparse autoencoders (SAEs) have emerged as a promising approach in language model interpretability, offering unsupervised extraction of sparse features. For interpretability methods to succeed, they must identify abstract features across domains, and these features can often manifest differently in each context. We examine this through "answerability" - a model's ability to recognize answerable questions. We extensively evaluate SAE feature generalization across diverse, partly self-constructed answerability datasets for Gemma 2 SAEs. Our analysis reveals that residual stream probes outperform SAE features within domains, but generalization performance differs sharply. SAE features show inconsistent out-of-domain transfer, with performance varying from almost random to outperforming residual stream probes. Overall, this demonstrates the need for robust evaluation methods and quantitative approaches to predict feature generalization in SAE-based interpretability.

Key facts. Cited by 8 papers in the map (data through 7 Oct 2026). Builds on 23 of its 39 references that are in the map. First posted on arXiv 27 Feb 2025.

Builds on

23 of 39 references.

References · 23

  1. 2024The Llama 3 Herd of Modelscited by 14,192
  2. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,928
  3. 2018Know What You Don't Know: Unanswerable Questions for SQuADcited by 1,538
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
  5. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,130
  6. 2022In-context Learning and Induction Headscited by 930
  7. 2023Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formattingcited by 778
  8. 2024Scaling and evaluating sparse autoencoderscited by 530
  9. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 412
  10. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 381
  11. 2023Finding Neurons in a Haystack: Case Studies with Sparse Probingcited by 328
  12. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 171
11 more
  1. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 148
  2. 2024Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Mapscited by 140
  3. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 109
  4. 2021Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factorscited by 106
  5. 2024Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Controlcited by 83
  6. 2023The Curious Case of Hallucinatory (Un)answerability: Finding Truths in the Hidden States of Over-Confident Large Language Modelscited by 54
  7. 2025Open Problems in Machine Unlearning for AI Safetycited by 49
  8. 2024Interpreting Attention Layer Outputs with Sparse Autoencoderscited by 43
  9. 2019Automatic Spanish Translation of the SQuAD Dataset for Multilingual Question Answeringcited by 35
  10. 2024Enhancing Neural Network Interpretability with Feature-Aligned Sparse Autoencoderscited by 25
  11. 2024Sparse Autoencoders Reveal Temporal Difference Learning in Large Language Modelscited by 15

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 5; 2026: 3.

8 papers.

  1. 2025No Answer Needed: Predicting LLM Answer Accuracy from Question-Only Linear Probescited by 35
  2. 2025Detecting (Un)answerability in Large Language Models with Linear Directionscited by 12
  3. 2025The Dead Salmons of AI Interpretabilitycited by 10
  4. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?cited by 10
  5. 2025Towards Combinatorial Interpretability of Neural Computationcited by 9
  6. 2026Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional Generalisationcited by 5
  7. 2025Latent Concept Disentanglement in Transformer-based Language Modelscited by 3
  8. 2026Sparse Reward Subsystem in Large Language Modelscited by 3