mlsanityData to 14 Oct 2026

Transferring Linear Features Across Language Models With Model Stitching

arXiv 2506.06609 ↗First posted 7 Jun 2025cs.CL cs.LGcited by 21

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

Abstract (arXiv)

In this work, we demonstrate that affine mappings between residual streams of language models is a cheap way to effectively transfer represented features between models. We apply this technique to transfer the weights of Sparse Autoencoders (SAEs) between models of different sizes to compare their representations. We find that small and large models learn similar representation spaces, which motivates training expensive components like SAEs on a smaller model and transferring to a larger model at a FLOPs savings. In particular, using a small-to-large transferred SAE as initialization can lead to 50% cheaper training runs when training SAEs on larger models. Next, we show that transferred probes and steering vectors can effectively recover ground truth performance. Finally, we dive deeper into feature-level transferability, finding that semantic and structural features transfer noticeably differently while specific classes of functional features have their roles faithfully mapped. Overall, our findings illustrate similarities and differences in the linear representation spaces of small and large models and demonstrate a method for improving the training efficiency of SAEs.

Key facts. Cited by 21 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 48 of its 67 references that are in the map. First posted on arXiv 7 Jun 2025.

Builds on

48 of 67 references.

Closest ancestors · 7

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. 2025Sparse Autoencoders Trained on the Same Data Learn Different Features28 Jan
  4. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders22 Sep
  5. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders19 Jul
  6. 2024Scaling and evaluating sparse autoencoders6 Jun
  7. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 41

  1. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
  2. 2020Scaling Laws for Neural Language Modelscited by 6,586
  3. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 3,104
  4. 2022Emergent Abilities of Large Language Modelscited by 2,512
  5. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  6. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,786
  7. 2019Similarity of Neural Network Representations Revisitedcited by 1,755
  8. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  9. 2023Instruction-Following Evaluation for Large Language Modelscited by 1,228
  10. 2022Editing Models with Task Arithmeticcited by 1,120
  11. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  12. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
29 more
  1. 2022Toy Models of Superpositioncited by 971
  2. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
  3. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  4. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  5. 2024The Platonic Representation Hypothesiscited by 382
  6. 2023Finding Neurons in a Haystack: Case Studies with Sparse Probingcited by 334
  7. 2023Language Models Represent Space and Timecited by 316
  8. 2023Function Vectors in Large Language Modelscited by 308
  9. 2023LEACE: Perfect linear concept erasure in closed formcited by 267
  10. 2014Understanding image representations by measuring their equivariance and equivalencecited by 243
  11. 2015Convergent Learning: Do different neural networks learn the same representations?cited by 220
  12. 2024Not All Language Model Features Are One-Dimensionally Linearcited by 212
  13. 2021Revisiting Model Stitching to Compare Neural Representationscited by 209
  14. 2025Open Problems in Mechanistic Interpretabilitycited by 179
  15. 2016Linear Algebraic Structure of Word Senses, with Applications to Polysemycited by 165
  16. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 160
  17. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
  18. 2024Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Modelscited by 115
  19. 2024Universal Neurons in GPT2 Language Modelscited by 102
  20. 2021Representation Topology Divergence: A Method for Comparing Neural Network Representationscited by 66
  21. 2024Confidence Regulation Neurons in Language Modelscited by 57
  22. 2024Spectral Filters, Dark Signals, and Attention Sinkscited by 56
  23. 2025Harnessing the Universal Geometry of Embeddingscited by 55
  24. 2025Do Language Models Use Their Depth Efficiently?cited by 50
  25. 2024Recurrent Neural Networks Learn to Store and Generate Sequences using Non-Linear Representationscited by 47
  26. 2024Decomposing The Dark Matter of Sparse Autoencoderscited by 45
  27. 2015Sparse Overcomplete Word Vector Representationscited by 44
  28. 2025Shared Global and Local Geometry of Language Model Embeddingscited by 36
  29. 2025Activation Space Interventions Can Be Transferred Between Large Language Modelscited by 19

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 21.

21 papers.

Sort:

By area:

  1. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  2. 2026Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reusecited by 15
  3. 2026Latent Cache Flow: Model-to-Model Communication Without Textcited by 8
  4. 2026A Negative Result on Cross-Model Activation Transfer in a Pythia Multi-Hop Settingcited by 4
  5. 2026When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMscited by 3
  6. 2026Frozen Memory Is Not Enough: Rethinking External Memory as Extractioncited by 3
  7. 2026Align Then Adapt: Label-Efficient Adapter Learning for Asymmetric Dense Retrievalcited by 2
  8. 2026Do Models Share Safety Representations? Cross-Model Steering for Safe Visual Generationcited by 2
  9. 2026Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanationscited by 2
  10. 2026One Adapter Pair per Model: A Universal Activation Interface for Language Modelscited by 2
  11. 2026ATLAS: Constitution-Conditioned Latent Geometry and Redistribution Across Language Models and Neural Perturbation Datacited by 1
  12. 2026Deep Minds and Shallow Probescited by 0
9 more
  1. 2026Polymorphism Is Rotation: Operational Mechanistic Interpretability from a Two-Layer Transformer to Pythia-70mcited by 0
  2. 2026Contrastive-Difference CKA Reveals Concept-Specific Structural Alignment Across Language Model Architecturescited by 0
  3. 2026How Far Do Simple Transformations Translate Across Text Embedding Models?cited by 0
  4. 2026SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalizationcited by 0
  5. 2026Approximate Homomorphisms and Convergent Representations in Transducerscited by 0
  6. 2026A Survey on the Linear Representation Hypothesiscited by 0
  7. 2026SuperAI: Merging Capabilities Across Heterogeneous Foundation Modelscited by 0
  8. 2026CacheLink: Functional KV Transport for Zero-Prefill Cross-Model Context Handoffcited by 0
  9. 2026Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Low-Rank Adapting Models for Sparse Autoencoders
  2. 2026When Can We Trust the Sparse Lens? A Certification Framework for SAE Faithfulness
  3. 2026Discovering Millions of Interpretable Features with Sparse Autoencoders
  4. 2025AdaptiveK: Complexity-Driven Sparse Autoencoders for Interpretable Language Model Representations
  5. 2025Interpreting the linear structure of vision-language model embedding spaces
  6. 2026Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models
  7. 2026Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility
  8. 2026Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression