mlsanityData to 7 Oct 2026

Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes

arXiv 2603.04426 ↗First posted 16 Feb 2026cs.LG cs.AIcited by 4

Not on the map. Why?

Abstract (arXiv)

Model diffing methods aim to identify how fine-tuning changes a model's internal representations. Crosscoders approach this by learning shared dictionaries of interpretable latent directions between base and fine-tuned models. However, existing formulations struggle with narrow fine-tuning, where behavioral changes are localized and asymmetric. We introduce Delta-Crosscoder, which combines BatchTopK sparsity with a delta-based loss prioritizing directions that change between models, plus an implicit contrastive signal from paired activations on matched inputs. Evaluated across 10 model organisms, including synthetic false facts, emergent misalignment, subliminal learning, and taboo word guessing (Gemma, LLaMA, Qwen; 1B-9B parameters), Delta-Crosscoder reliably isolates latent directions causally responsible for fine-tuned behaviors and enables effective mitigation, outperforming SAE-based baselines, while matching the Non-SAE-based. Our results demonstrate that crosscoders remain a powerful tool for model diffing.

Key facts. Cited by 4 papers in the map (data through 7 Oct 2026). Builds on 30 of its 42 references that are in the map. First posted on arXiv 16 Feb 2026.

Builds on

30 of 42 references.

References · 30

  1. 2024The Llama 3 Herd of Modelscited by 14,192
  2. 2025Qwen3 Technical Reportcited by 8,195
  3. 2025Gemma 3 Technical Reportcited by 1,686
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
  5. 2024The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scalecited by 1,104
  6. 2022Toy Models of Superpositioncited by 928
  7. 2023LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Datasetcited by 449
  8. 2023Finding Neurons in a Haystack: Case Studies with Sparse Probingcited by 328
  9. 2024Alignment faking in large language modelscited by 316
  10. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 303
  11. 2025Layer by Layer: Uncovering Hidden Representations in Language Modelscited by 293
  12. 2024Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelscited by 225
18 more
  1. 2020What Happens To BERT Embeddings During Fine-tuning?cited by 113
  2. 2025Natural Emergent Misalignment from Reward Hacking in Production RLcited by 111
  3. 2023Mechanistically analyzing the effects of fine-tuning on procedurally defined taskscited by 110
  4. 2023HuatuoGPT-II, One-stage Training for Medical Adaption of LLMscited by 91
  5. 2025Subliminal Learning: Language models transmit behavioral traits via hidden signals in datacited by 89
  6. 2025Model Organisms for Emergent Misalignmentcited by 84
  7. 2025Persona Features Control Emergent Misalignmentcited by 76
  8. 2025Convergent Linear Representations of Emergent Misalignmentcited by 66
  9. 2023From Language Modeling to Instruction Following: Understanding the Behavior Shift in LLMs after Instruction Tuningcited by 56
  10. 2025Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainerscited by 41
  11. 2025Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuningcited by 34
  12. 2025Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuningcited by 31
  13. 2025Narrow Finetuning Leaves Clearly Readable Traces in Activation Differencescited by 24
  14. 2025Eliciting Secret Knowledge from Language Modelscited by 23
  15. 2025Believe It or Not: How Deeply do LLMs Believe Implanted Facts?cited by 19
  16. 2024On Domain-Adaptive Post-Training for Multimodal Large Language Modelscited by 13
  17. 2026Cross-Architecture Model Diffing with Crosscoders: Unsupervised Discovery of Differences Between LLMscited by 12
  18. 2025Reviving Your MNEME: Predicting The Side Effects of LLM Unlearning and Fine-Tuning via Sparse Model Diffingcited by 8

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 4.

4 papers.

  1. 2026Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoningcited by 0
  2. 2026Localizing RL-Induced Tool Use to a Single Crosscoder Featurecited by 0
  3. 2026Introspecting Alignment Shifts Beyond Behaviors Implanted Through Fine-Tuningcited by 0
  4. 2026From Interpretability Methods to Interpretable Modelscited by 0