mlsanityData to 14 Oct 2026

Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures

arXiv 2410.06672 ↗First posted 9 Oct 2024cs.CLcited by 31

On the map: State-space sequence models (S4, Mamba) → Core SSM architectures for language

also in 1 more
Abstract (arXiv)

The hypothesis of Universality in interpretability suggests that different neural networks may converge to implement similar algorithms on similar tasks. In this work, we investigate two mainstream architectures for language modeling, namely Transformers and Mambas, to explore the extent of their mechanistic similarity. We propose to use Sparse Autoencoders (SAEs) to isolate interpretable features from these models and show that most features are similar in these two models. We also validate the correlation between feature similarity and Universality. We then delve into the circuit-level analysis of Mamba models and find that the induction circuits in Mamba are structurally analogous to those in Transformers. We also identify a nuanced difference we call Off-by-One motif: The information of one token is written into the SSM state in its next position. Whilst interaction between tokens in Transformers does not exhibit such trend.

Key facts. Cited by 31 papers in the map (data through 14 Oct 2026). Placed in Core SSM architectures for language, one of 10 directions in State-space sequence models (S4, Mamba). Builds on 35 of its 46 references that are in the map. First posted on arXiv 9 Oct 2024.

Builds on

35 of 46 references.

Closest ancestors · 4

Same direction, newest first

  1. 2024Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality31 May
  2. 2024Locating and Editing Factual Associations in Mamba4 Apr
  3. 2024The Hidden Attention of Mamba Models3 Mar
  4. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spaces1 Dec

Other references · 31

  1. 2017Attention Is All You Needcited by 60,361
  2. 2020Language Models are Few-Shot Learnerscited by 32,419
  3. 2022Training language models to follow instructions with human feedbackcited by 15,971
  4. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  5. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
  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. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  9. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  10. 2022Toy Models of Superpositioncited by 971
  11. 2022In-context Learning and Induction Headscited by 964
  12. 2023RWKV: Reinventing RNNs for the Transformer Eracited by 827
19 more
  1. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  2. 2023Retentive Network: A Successor to Transformer for Large Language Modelscited by 613
  3. 2022Hungry Hungry Hippos: Towards Language Modeling with State Space Modelscited by 499
  4. 2024The Platonic Representation Hypothesiscited by 382
  5. 2023Function Vectors in Large Language Modelscited by 308
  6. 2015Convergent Learning: Do different neural networks learn the same representations?cited by 220
  7. 2022Transformer Quality in Linear Timecited by 191
  8. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 179
  9. 2023A Toy Model of Universality: Reverse Engineering How Networks Learn Group Operationscited by 144
  10. 2023Similarity of Neural Network Models: A Survey of Functional and Representational Measurescited by 131
  11. 2018Identifying and Controlling Important Neurons in Neural Machine Translationcited by 111
  12. 2024Universal Neurons in GPT2 Language Modelscited by 102
  13. 2018Towards Understanding Learning Representations: To What Extent Do Different Neural Networks Learn the Same Representationcited by 62
  14. 2018Geometry Score: A Method For Comparing Generative Adversarial Networkscited by 57
  15. 2024Interpreting the Second-Order Effects of Neurons in CLIPcited by 36
  16. 2021Grounding Representation Similarity with Statistical Testingcited by 29
  17. 2024Dictionary Learning Improves Patch-Free Circuit Discovery in Mechanistic Interpretability: A Case Study on Othello-GPTcited by 29
  18. 2024Automatically Identifying Local and Global Circuits with Linear Computation Graphscited by 24
  19. 2022GULP: a prediction-based metric between representationscited by 17

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2024: 1; 2025: 13; 2026: 17.

31 papers.

Sort:

By area:

  1. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 143
  2. 2025Analyze Feature Flow to Enhance Interpretation and Steering in Language Modelscited by 13
  3. 2025Layers at Similar Depths Generate Similar Activations Across LLM Architecturescited by 12
  4. 2025Representation Potentials of Foundation Models for Multimodal Alignment: A Surveycited by 11
  5. 2025Towards Understanding the Nature of Attention with Low-Rank Sparse Decompositioncited by 9
  6. 2025Comprehension Without Competence: Architectural Limits of LLMs in Symbolic Computation and Reasoningcited by 9
  7. 2025Evolution of Concepts in Language Model Pre-Trainingcited by 8
  8. 2025Dimensional Collapse in Transformer Attention Outputs: A Challenge for Sparse Dictionary Learningcited by 6
  9. 2025How a Bilingual LM Becomes Bilingual: Tracing Internal Representations with Sparse Autoencoderscited by 5
  10. 2026Measuring Affinity between Attention-Head Weight Subspaces via the Projection Kernelcited by 5
  11. 2025Superposition disentanglement of neural representations reveals hidden alignmentcited by 4
  12. 2026The Indra Representation Hypothesis for Multimodal Alignmentcited by 4
19 more
  1. 2025Model Alignment Searchcited by 3
  2. 2026Interpreting and Steering State-Space Models via Activation Subspace Bottleneckscited by 3
  3. 2026Differentiable Faithfulness Alignment for Cross-Model Circuit Transfercited by 3
  4. 2026Cross-Architecture Steering Transfer in Language Models: A Systematic Empirical Studycited by 2
  5. 2025Explaining How Visual, Textual and Multimodal Encoders Share Conceptscited by 1
  6. 2026WriteSAE: Sparse Autoencoders for Recurrent Statecited by 1
  7. 2026Interpretability Transfer from Language to Vision via Sparse Autoencoderscited by 1
  8. 2026Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoderscited by 1
  9. 2025Universal Neurons in GPT-2: Emergence, Persistence, and Functional Impactcited by 0
  10. 2025The Deleuzian Representation Hypothesiscited by 0
  11. 2026Refit the Probe: Single-Direction Ablation Is Not a Necessity Testcited by 0
  12. 2026Contrastive-Difference CKA Reveals Concept-Specific Structural Alignment Across Language Model Architecturescited by 0
  13. 2026What, Where, and How: Disentangling the Roles of Task, Language, and Model in Code Model Representationscited by 0
  14. 2026Mapping and Measuring the Behavioral Evolution of Large Language Modelscited by 0
  15. 2026SharedSAE: One Feature Dictionary Across Language Modelscited by 0
  16. 2026Locating and Steering Refusal Beyond Attentioncited by 0
  17. 2026Global Divergence, Local Convergence: Representation Geometry in SSMs and Transformerscited by 0
  18. 2026Comparing Latent Concept Formation in State Space Models and Transformers via Sparse Autoencoderscited by 0
  19. 2026What a Cross-Model Fixed-Point Census Can and Cannot Arbitrate About Repetitioncited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoders
  2. 2024A Practical Review of Mechanistic Interpretability for Transformer-Based Language Models
  3. 2025Semantic Convergence: Investigating Shared Representations Across Scaled LLMs
  4. 2024An Empirical Study of Mamba-based Language Models
  5. 2026Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability
  6. 2024Transcoders Find Interpretable LLM Feature Circuits
  7. 2026Mechanistic Interpretability for Large Language Model Alignment: Progress, Challenges, and Future Directions
  8. 2024Circuit Compositions: Exploring Modular Structures in Transformer-Based Language Models