mlsanityData to 14 Oct 2026

Are Sparse Autoencoder Benchmarks Reliable?

arXiv 2605.18229 ↗First posted 18 May 2026cs.LG cs.AIcited by 4

On the map: Mechanistic interpretability → SAE features and their applications

Abstract (arXiv)

Sparse autoencoders (SAEs) are a core interpretability tool for large language models, and progress on SAE architectures depends on benchmarks that reliably distinguish better SAEs from worse ones. We audit the SAE quality metrics in SAEBench, the de-facto standard SAE evaluation suite, through three complementary lenses: reseed noise on a fixed SAE, ground-truth correlation on synthetic SAEs, and discriminability across training trajectories. We find that two of these metrics, Targeted Probe Perturbation (TPP) and Spurious Correlation Removal (SCR), fail multiple lenses at their canonical settings and should not be used to evaluate SAEs. The other metrics show higher reseed noise and lower discriminability than the field assumes. The sae-probes variant of k-sparse probing is the most reliable metric we tested, but even sae-probes struggles to separate variants of the same SAE architecture. Our results show the field needs better SAE benchmarks.

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Placed in SAE features and their applications, one of 13 directions in Mechanistic interpretability. Builds on 26 of its 31 references that are in the map. First posted on arXiv 18 May 2026.

Builds on

26 of 31 references.

Closest ancestors · 9

Same direction, newest first

  1. 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability12 Mar
  2. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probing23 Feb
  3. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders28 Jan
  4. 2025Sparse Autoencoders Trained on the Same Data Learn Different Features28 Jan
  5. 2024BatchTopK Sparse Autoencoders9 Dec
  6. 2024Automatically Interpreting Millions of Features in Large Language Models17 Oct
  7. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders22 Sep
  8. 2024Measuring Progress in Dictionary Learning for Language Model Interpretability with Board Game Models31 Jul
  9. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 17

  1. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
  2. 2022Toy Models of Superpositioncited by 971
  3. 2017Deep Reinforcement Learning that Matterscited by 754
  4. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  5. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  6. 2023Finding Neurons in a Haystack: Case Studies with Sparse Probingcited by 334
  7. 2021What Will it Take to Fix Benchmarking in Natural Language Understanding?cited by 203
  8. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
  9. 2020With Little Power Comes Great Responsibilitycited by 142
  10. 2021The Benchmark Lotterycited by 136
  11. 2024Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Controlcited by 86
  12. 2024RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representationscited by 79
5 more
  1. 2017Reporting Score Distributions Makes a Difference: Performance Study of LSTM-networks for Sequence Taggingcited by 73
  2. 2024Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoderscited by 18
  3. 2025Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEscited by 14
  4. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?cited by 14
  5. 2026SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Datacited by 11

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 4.

4 papers.

  1. 2026ICA Lens: Interpreting Language Models Without Training Another Dictionarycited by 4
  2. 2026SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?cited by 0
  3. 2026Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Trainingcited by 0
  4. 2026SFT Leaves a Reproducible Signature in Activation Geometry: Calibrating Post-Training Interpretability Against Run-to-Run Noisecited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks
  2. 2026Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility
  3. 2026Discovering Millions of Interpretable Features with Sparse Autoencoders
  4. 2025Revisiting End-To-End Sparse Autoencoder Training: A Short Finetune Is All You Need
  5. 2024Improving Dictionary Learning with Gated Sparse Autoencoders
  6. 2026When Can We Trust the Sparse Lens? A Certification Framework for SAE Faithfulness
  7. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders
  8. 2024SAGE: Scalable Ground Truth Evaluations for Large Sparse Autoencoders