mlsanityData to 14 Oct 2026

Interpreting and steering diffusion models

27 papers0 this weekMedium confidence in this grouping: mostly one topic, with blurry edgesMedium confidence in this grouping: mostly one topic, with blurry edges · About

Keywords: diffusion, concept, erasure, sparse, text-to-image

Key facts. Interpreting and steering diffusion models: 27 papers on arXiv since 2022 (data through 14 Oct 2026). 11 of its 27 mapped papers (41%) cite “SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders” (2025, arXiv 2501.18052). 4 new papers 24 Sep–14 Oct.

What’s new

24 Sep – 14 Oct 2026

  1. Earlier · 4
  2. Beyond the Linear Representation Hypothesis: Non-Linear Activation Steering in Text-to-Image Models also fits: Interpreting vision and multimodal models
  3. RASteer: Retain-Aware Activation Steering for Concept Erasure in Diffusion Models
  4. D-Scope: Decomposing and Steering Diffusion Transformers with Sparse Autoencoders
  5. SharedSAE: a unified sparse dictionary across u-net blocks and denoising time for diffusion models

Foundations

Most cited by papers here

  1. SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders2025cited by 11 here
  2. One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models2024cited by 7 here
  3. textitRevelio: Interpreting and leveraging semantic information in diffusion models2024cited by 5 here
  4. TIDE : Temporal-Aware Sparse Autoencoders for Interpretable Diffusion Transformers in Image Generation2025cited by 4 here
  5. Concept Steerers: Leveraging K-Sparse Autoencoders for Test-Time Controllable Generations2025cited by 3 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Extracting Latent Steering Vectors from Pretrained Language Modelscited by 2 here
  2. 2023Steering Language Models With Activation Engineeringcited by 4 here
  3. 2023Steering Llama 2 via Contrastive Activation Additioncited by 2 here
  4. 2024Interpreting Attention Layer Outputs with Sparse Autoencoderscited by 3 here
  5. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 2 here
  6. 2024One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Modelscited by 7 here
  7. 2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 11 here
17 more it builds on, not in this chain
  1. 2022Toy Models of Superpositioncited by 4 here
  2. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 20 here
  3. 2024Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelscited by 2 here
  4. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 3 here
  5. 2024The Missing Curve Detectors of InceptionV1: Applying Sparse Autoencoders to InceptionV1 Early Visioncited by 4 here
  6. 2024Scaling and evaluating sparse autoencoderscited by 10 here
  7. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 2 here
  8. 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 2 here
  9. 2024Interpreting and Editing Vision-Language Representations to Mitigate Hallucinationscited by 3 here
  10. 2024An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generationcited by 3 here
  11. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 3 here
  12. 2024textitRevelio: Interpreting and leveraging semantic information in diffusion modelscited by 5 here
  13. 2024BatchTopK Sparse Autoencoderscited by 4 here
  14. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 4 here
  15. 2025Concept Steerers: Leveraging K-Sparse Autoencoders for Test-Time Controllable Generationscited by 3 here
  16. 2025TIDE : Temporal-Aware Sparse Autoencoders for Interpretable Diffusion Transformers in Image Generationcited by 4 here
  17. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 3 here

The 24 most cited of 64.

Browse

22 of 27, ranked 6–27

  1. 6Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning20252
  2. 7SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders20252
  3. 8A Single Neuron Works: Precise Concept Erasure in Text-to-Image Diffusion Models20252
  4. 9OrthoEraser: Coupled-Neuron Orthogonal Projection for Concept Erasure20261
  5. 10Emergence and Evolution of Interpretable Concepts in Diffusion Models20251
  6. 11CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models20260
  7. 12DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders20260
  8. 13Learning a Generative Meta-Model of LLM Activations20260
  9. 14PolypSteer: Counterfactual Endoscopic Synthesis via Training-Free Activation Steering20260
  10. 15SHIFT: Steering Hidden Intermediates in Flow Transformers20260
All 22
  1. 16Deep Dreams Are Made of This: Visualizing Monosemantic Features in Diffusion Models20260
  2. 17Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models20260
  3. 18DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing20260
  4. 19Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models20260
  5. 20PEAK: Precise and Persistent Concept Erasure via k-Sparse Autoencoders20260
  6. 21EraseSAE: Surgical Concept Erasure in Text-to-Video Diffusion Models via Sparse Autoencoders20260
  7. 22Sparse Autoencoder as a Zero-Shot Classifier for Concept Erasing in Text-to-Image Diffusion Models20250
  8. 23Decoding Vision Transformers: the Diffusion Steering Lens20250
  9. 24Model-Agnostic Gender Bias Control for Text-to-Image Generation via Sparse Autoencoder20250
  10. 25SAEdit: Token-level control for continuous image editing via Sparse AutoEncoder20250
  11. 26LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfer20250
  12. 27Schr"odinger's Bat: Diffusion Models Sometimes Generate Polysemous Words in Superposition20220