Elamur Research⌕
☰
home / papers / 2026 / From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?

From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?

Paper · 2026 · Aaron Mueller, Andrew Lee, Shruti Joshi et al.

Summary

Interpretability research usually judges featurizers such as sparse autoencoders (SAEs) and probes one concept at a time, implicitly assuming that concepts are independent. The authors ask whether representations that look disentangled actually let each concept be controlled on its own. They build a synthetic English dataset with a probabilistic context-free grammar (PCFG). Every sentence carries four ground-truth labels: voice, tense, sentiment and domain. The co-occurrence of chosen concept values can be raised through upsampling. Several SAE variants and sparse linear probes are applied to frozen Pythia-70M and Gemma-2-2B activations. They are scored with correlational metrics from causal representation learning (MCC, DCI-ES) and then tested in steering experiments.

On correlational metrics the featurizers do well. Individual features mostly track a single concept, although one concept is often spread across many features. However, steering the strongest feature for one concept frequently shifts other concepts as well, even though feature pairs show almost no interaction effects. The lesson is that neither correlational scores nor non-overlapping subspaces guarantee selective control, so multi-concept, interventional evaluation is necessary.

Key points

  • Dataset: PCFG-generated, with 382,459 training and 1,007 test sentences.
    • Binary concepts: voice (active/passive) and tense (present/past).
    • Ordinal three-way concept: sentiment (positive/neutral/negative).
    • Unordered four-way concept: domain (news/science/fantasy/other).
    • Correlations between concept values, such as positive sentiment with science, are adjustable.
  • Featurizers compared: ReLU SAEs (public ones from Marks et al. for Pythia), JumpReLU SAEs (Lieberum et al. for Gemma), Top-K, SpADE, and SSAEs trained on activation differences. Supervised sparse probes serve as a baseline.
  • Correlational results: Identifiability failures arise mainly from feature splitting, where many features cover one concept, rather than from single features mixing concepts. Featurizers also improve on the model's raw neuron basis.
  • Steering results: The authors propose metrics for steering entanglement. Selective steerability and feature disjointness (non-interacting effects) turn out to be distinct, and neither predicts the other. Non-overlapping circuits or subspaces may therefore be a weak proxy for mechanistic independence.
  • Limitations and availability: The study covers only four categorical concepts on a narrow PCFG distribution. Data and code are on GitHub (aaronmueller/IdentifiableLanguage).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Harvard UniversityCambridge · United States112 works · 42,023 citations🔬Mila - Quebec Artificial Intelligence InstituteMontreal · Canada54 works · 13,565 citations🎓University of TübingenTübingen · Germany75 works · 41,081 citations🎓Boston UniversityBoston · United States26 works · 4,022 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Interpretability · 130 works

Cite

Cite as Aaron Mueller, Andrew Lee, Shruti Joshi, Ekdeep Singh Lubana, Dhanya Sridhar and Patrik Reizinger. 2026. From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.782

@inproceedings{mueller2026isolation,
  title = {{From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?}},
  author = {Aaron Mueller, Andrew Lee, Shruti Joshi, Ekdeep Singh Lubana, Dhanya Sridhar and Patrik Reizinger},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.782},
  url = {https://doi.org/10.18653/v1/2026.acl-long.782}
}

References

73 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Kartik Ahuja, Jason S Hartford, and Yoshua Bengio. 2022. Weakly supervised representation learning with sparse perturbations. Advances in Neural Information Processing Systems, 35:15516–15528. find
  2. Kartik Ahuja, Divyat Mahajan, Yixin Wang, and Yoshua Bengio. 2023a. Interventional causal representation learning. In International Conference on Machine Learning, pages 372–407. PMLR. find
  3. Kartik Ahuja, Amin Mansouri, and Yixin Wang. 2023b. Multi-Domain Causal Representation Learning via Weak Distributional Invariances. arXiv preprint. ArXiv:2310.02854 [cs, stat]. — [arXiv:2310.02854](https://arxiv.org/abs/2310.02854) find
  4. Dana Arad, Aaron Mueller, and Yonatan Belinkov. 2025. SAEs are good for steering – if you select the right features. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 10241–10259, Suzhou, China. Association for Computational Linguistics.
  5. Andy Arditi, Oscar Balcells Obeso, Aaquib Syed, Daniel Paleka, Nina Rimsky, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.
  6. Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar Van Der Wal. 2023. Pythia: a suite for analyzing large language models across trainin… find
  7. Taylor L. Booth and Richarda . Thompson. 1973. Applying probability measures to abstract languages. IEEE Transactions on Computers, C-22:442–450. find
  8. Jack Brady, Julius von Kügelgen, Sebastien Lachapelle, Simon Buchholz, Thomas Kipf, and Wieland Brendel. 2025. Interaction asymmetry: A general principle for learning composable abstractions. In The Thirteenth International Conference on Learning Representations. find
  9. Johann Brehmer, Pim de Haan, Phillip Lippe, and Taco Cohen. 2022. Weakly supervised causal representation learning. arXiv preprint. ArXiv:2203.16437 [cs, stat]. — [arXiv:2203.16437](https://arxiv.org/abs/2203.16437) find
  10. Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, and 6 others. 2023. To… find
  11. Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Schölkopf, and Pradeep Ravikumar. 2023. Learning linear causal representations from interventions under general nonlinear mixing. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023). find
  12. Valérie Costa, Thomas Fel, Ekdeep Singh Lubana, Bahareh Tolooshams, and Demba E. Ba. 2025. From flat to hierarchical: Extracting sparse representations with matching pursuit. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. find
  13. George Darmois. 1951. Analyse des liaisons de prob-abilité. In Proc. find
  14. Int. Stat. Conferences 1947, page 231. find
  15. Cian Eastwood, Armin Kekic, and Andrei Liviu Nicolicioiu. 2022. On the DCI Framework for Evaluating Disentangled Representations: Extensions and Connections to Identifiability. page 8. find
  16. Cian Eastwood, Andrei Liviu Nicolicioiu, Julius Von Kügelgen, Armin Kekić, Frederik Träuble, Andrea Dittadi, and Bernhard Schölkopf. 2023. DCI-ES: An extended disentanglement framework with connections to identifiability. In The Eleventh International Conference on Learning Representations. find
  17. Cian Eastwood and Christopher K. I. Williams. 2018. A framework for the quantitative evaluation of disentangled representations. In International Conference on Learning Representations. find
  18. Joshua Engels, Eric J Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. 2025. Not all language model features are one-dimensionally linear. In The Thirteenth International Conference on Learning Representations. find
  19. Jose Gallego-Posada and Juan Ramirez. 2022. Cooper: a toolkit for Lagrangian-based constrained optimization. https://github.com/cooper-org/cooper. ↗
  20. Leo Gao, Tom Dupre la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2025. Scaling and evaluating sparse autoencoders. In The Thirteenth International Conference on Learning Representations. find
  21. Atticus Geiger, Duligur Ibeling, Amir Zur, Maheep Chaudhary, Sonakshi Chauhan, Jing Huang, Aryaman Arora, Zhengxuan Wu, Noah Goodman, Christopher Potts, and Thomas Icard. 2025. Causal abstraction: A theoretical foundation for mechanistic interpretability. Journal of Machine Learning Research, 26(83)… find
  22. Gauthier Gidel, Hugo Berard, Gaëtan Vignoud, Pascal Vincent, and Simon Lacoste-Julien. 2020. A variational inequality perspective on generative adversarial networks. Preprint, arXiv:1802.10551. — [arXiv:1802.10551](https://arxiv.org/abs/1802.10551) find
  23. Luigi Gresele, Paul K. Rubenstein, Arash Mehrjou, Francesco Locatello, and Bernhard Schölkopf. 2019. The Incomplete Rosetta Stone Problem: Identifiability Results for Multi-View Nonlinear ICA. arXiv:1905.06642 [cs, stat]. ArXiv: 1905.06642. — [arXiv:1905.06642](https://arxiv.org/abs/1905.06642) find
  24. Luigi Gresele, Julius von Kügelgen, Vincent Stimper, Bernhard Schölkopf, and Michel Besserve. 2021. Independent mechanism analysis, a new concept? arXiv:2106.05200 [cs, stat]. ArXiv: 2106.05200. — [arXiv:2106.05200](https://arxiv.org/abs/2106.05200) find
  25. Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsimas. 2023. Finding neurons in a haystack: Case studies with sparse probing. Transactions on Machine Learning Research. find
  26. Irina Higgins, David Amos, David Pfau, Sebastien Racaniere, Loic Matthey, Danilo Rezende, and Alexander Lerchner. 2018. Towards a Definition of Disentangled Representations. arXiv:1812.02230 [cs, stat]. ArXiv: 1812.02230. — [arXiv:1812.02230](https://arxiv.org/abs/1812.02230) find
  27. Sai Sumedh R. Hindupur, Ekdeep Singh Lubana, Thomas Fel, and Demba Ba. 2025. Projecting assumptions: The duality between sparse autoencoders and concept geometry. Preprint, arXiv:2503.01822. — [arXiv:2503.01822](https://arxiv.org/abs/2503.01822) find
  28. Kyle Hsu, Will Dorrell, James C. R. Whittington, Jiajun Wu, and Chelsea Finn. 2023. Disentanglement via Latent Quantization. arXiv preprint. ArXiv:2305.18378 [cs, stat]. — [arXiv:2305.18378](https://arxiv.org/abs/2305.18378) find
  29. Jing Huang, Zhengxuan Wu, Christopher Potts, Mor Geva, and Atticus Geiger. 2024. RAVEL: Evaluating interpretability methods on disentangling language model representations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8669–… find
  30. Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. 2024. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations. find
  31. Aapo Hyvarinen and Hiroshi Morioka. 2016. Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA. arXiv:1605.06336 [cs, stat]. ArXiv: 1605.06336. — [arXiv:1605.06336](https://arxiv.org/abs/1605.06336) find
  32. Aapo Hyvarinen and Hiroshi Morioka. 2017. Nonlinear ICA of Temporally Dependent Stationary Sources. In Artificial Intelligence and Statistics, pages 460–469. PMLR. ISSN: 2640-3498. find
  33. Aapo Hyvarinen, Hiroaki Sasaki, and Richard E. Turner. 2019. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. arXiv:1805.08651 [cs, stat]. ArXiv: 1805.08651. — [arXiv:1805.08651](https://arxiv.org/abs/1805.08651) find
  34. Aapo Hyvärinen and Petteri Pajunen. 1999. Nonlinear independent component analysis: Existence and uniqueness results. Neural Networks, 12(3):429–439. find
  35. Shruti Joshi, Andrea Dittadi, Sébastien Lachapelle, and Dhanya Sridhar. 2025. Identifiable Steering via Sparse Autoencoding of Multi-Concept Shifts. arXiv preprint. ArXiv:2502.12179 [cs]. — [arXiv:2502.12179](https://arxiv.org/abs/2502.12179) find
  36. Shruti Joshi, Théo Saulus, Wieland Brendel, Philippe Brouillard, Dhanya Sridhar, and Patrik Reizinger. 2026. Who guards the guardians? The challenges of evaluating identifiability of learned representations. Preprint, arXiv:2602.24278. — [arXiv:2602.24278](https://arxiv.org/abs/2602.24278) find
  37. Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. 2020a. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. In International Conference on Artificial Intelligence and Statistics, pages 2207–2217. PMLR. ISSN: 2640-3498. find
  38. Ilyes Khemakhem, Ricardo Pio Monti, Diederik P. Kingma, and Aapo Hyvärinen. 2020b. ICE-BeeM: Identifiable Conditional Energy-Based Deep Models Based on Nonlinear ICA. arXiv:2002.11537 [cs, stat]. ArXiv: 2002.11537. — [arXiv:2002.11537](https://arxiv.org/abs/2002.11537) find
  39. Sebastien Lachapelle, Tristan Deleu, Divyat Mahajan, Ioannis Mitliagkas, Yoshua Bengio, Simon Lacoste-Julien, and Quentin Bertrand. 2023a. Synergies between Disentanglement and Sparsity: Generalization and Identifiability in Multi-Task Learning. In Proceedings of the 40th International Conference on… find
  40. Sébastien Lachapelle, Pau Rodrı́guez López, Yash Sharma, Katie Everett, Rémi Le Priol, Alexandre Lacoste, and Simon Lacoste-Julien. 2021. Disentanglement via Mechanism Sparsity Regularization: A New Principle for Nonlinear ICA. arXiv:2107.10098 [cs, stat]. ArXiv: 2107.10098. — [arXiv:2107.10098](https://arxiv.org/abs/2107.10098) find
  41. Sébastien Lachapelle, Divyat Mahajan, Ioannis Mitliagkas, and Simon Lacoste-Julien. 2023b. Additive Decoders for Latent Variables Identification and Cartesian-Product Extrapolation. arXiv preprint. ArXiv:2307.02598 [cs, stat]. — [arXiv:2307.02598](https://arxiv.org/abs/2307.02598) find
  42. Elliot Layne, Jason Hartford, Sébastien Lachapelle, Mathieu Blanchette, and Dhanya Sridhar. 2025. Sparsity regularization via tree-structured environments for disentangled representations. Transactions on Machine Learning Research. find
  43. Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, Janos Kramar, Anca Dragan, Rohin Shah, and Neel Nanda. 2024. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2. In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Inte… find
  44. Phillip Lippe, Sara Magliacane, Sindy Löwe, Yuki M. Asano, Taco Cohen, and Efstratios Gavves. 2022. CITRIS: Causal Identifiability from Temporal Intervened Sequences. arXiv preprint. Number: arXiv:2202.03169 arXiv:2202.03169 [cs, stat]. — [arXiv:2202.03169](https://arxiv.org/abs/2202.03169) find
  45. Phillip Lippe, Sara Magliacane, Sindy Löwe, Yuki M. Asano, Taco Cohen, and Efstratios Gavves. 2023. BISCUIT: Causal Representation Learning from Binary Interactions. arXiv preprint. ArXiv:2306.09643 [cs, stat]. — [arXiv:2306.09643](https://arxiv.org/abs/2306.09643) find
  46. Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. 2019. Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations. In International Conference on Machine Learning, pages 4114–4124. PMLR. ISSN: 264… find
  47. Francesco Locatello, Ben Poole, Gunnar Rätsch, Bernhard Schölkopf, Olivier Bachem, and Michael Tschannen. 2020a. Weakly-Supervised Disentanglement Without Compromises. arXiv:2002.02886 [cs, stat]. ArXiv: 2002.02886. — [arXiv:2002.02886](https://arxiv.org/abs/2002.02886) find
  48. Francesco Locatello, Michael Tschannen, Stefan Bauer, Gunnar Rätsch, Bernhard Schölkopf, and Olivier Bachem. 2020b. Disentangling Factors of Variation Using Few Labels. arXiv:1905.01258 [cs, stat]. ArXiv: 1905.01258. — [arXiv:1905.01258](https://arxiv.org/abs/1905.01258) find
  49. Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. 2025. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. In The Thirteenth International Conference on Learning Representations. find
  50. Samuel Marks and Max Tegmark. 2024. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. In First Conference on Language Modeling.
  51. Gemma E. Moran, Dhanya Sridhar, Yixin Wang, and David M. Blei. 2022. Identifiable Deep Generative Models via Sparse Decoding. Technical Report arXiv:2110.10804, arXiv. ArXiv:2110.10804 [cs, stat] type: article. find
  52. Aaron Mueller, Jannik Brinkmann, Millicent Li, Samuel Marks, Koyena Pal, Nikhil Prakash, Can Rager, Aruna Sankaranarayanan, Arnab Sen Sharma, Jiuding Sun, Eric Todd, David Bau, and Yonatan Belinkov. 2025a. The quest for the right mediator: Surveying mechanistic interpretability for nlp through the l…
  53. Aaron Mueller, Atticus Geiger, Sarah Wiegreffe, Dana Arad, Iván Arcuschin, Adam Belfki, Yik Siu Chan, Jaden Fried Fiotto-Kaufman, Tal Haklay, Michael Hanna, Jing Huang, Rohan Gupta, Yaniv Nikankin, Hadas Orgad, Nikhil Prakash, Anja Reusch, Aruna Sankaranarayanan, Shun Shao, Alessandro Stolfo, and 4… find
  54. Bruno A. Olshausen and David J. Field. 1997. Sparse coding with an overcomplete basis set: A strategy employed by v1? Vision Research, 37(23):3311– 3325. find
  55. Judea Pearl. 2009. Causal inference in statistics: An overview. Statistics Surveys, 3(none). find
  56. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:282… find
  57. Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. 2018. Elements of causal inference: foundations and learning algorithms. Journal of Statistical Computation and Simulation, 88(16):3248–3248. find
  58. Patrik Reizinger, Luigi Gresele, Jack Brady, Julius von Kügelgen, Dominik Zietlow, Bernhard Schölkopf, Georg Martius, Wieland Brendel, and Michel Besserve. 2023a. Embrace the Gap: VAEs Perform Independent Mechanism Analysis. arXiv preprint. ArXiv:2206.02416 [cs, stat]. — [arXiv:2206.02416](https://arxiv.org/abs/2206.02416) find
  59. Patrik Reizinger, Siyuan Guo, Ferenc Huszár, Bernhard Schölkopf, and Wieland Brendel. 2024. Identifiable Exchangeable Mechanisms for Causal Structure and Representation Learning. find
  60. Patrik Reizinger, Yash Sharma, Matthias Bethge, Bernhard Schölkopf, Ferenc Huszár, and Wieland Brendel. 2023b. Jacobian-based Causal Discovery with Nonlinear ICA. Transactions on Machine Learning Research. find
  61. Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. 2021. Towards Causal Representation Learning. arXiv:2102.11107 [cs]. ArXiv: 2102.11107 version: 1. find
  62. Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. 2014. Deep inside convolutional networks: Visualising image classification models and saliency maps. Preprint, arXiv:1312.6034.
  63. Nishant Subramani, Nivedita Suresh, and Matthew Peters. 2022. Extracting latent steering vectors from pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566–581, Dublin, Ireland. Association for Computational Linguistics.
  64. Raphael Suter, Djordje Miladinovic, Bernhard Schölkopf, and Stefan Bauer. 2019. Robustly disentangled causal mechanisms: Validating deep representations for interventional robustness. In International Conference on Machine Learning, pages 6056–6065. PMLR. find
  65. Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose… find
  66. Julius von Kügelgen. 2024. Identifiable Causal Representation Learning: Unsupervised, Multi-View, and Multi-Environment. find
  67. Julius von Kügelgen, Michel Besserve, Liang Wendong, Luigi Gresele, Armin Kekić, Elias Bareinboim, David M. Blei, and Bernhard Schölkopf. 2023. Nonparametric Identifiability of Causal Representations from Unknown Interventions. arXiv preprint. ArXiv:2306.00542 [cs, stat]. find
  68. Julius von Kügelgen, Yash Sharma, Luigi Gresele, Wieland Brendel, Bernhard Schölkopf, Michel Besserve, and Francesco Locatello. 2021. Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style. ArXiv: 2106.04619. — [arXiv:2106.04619](https://arxiv.org/abs/2106.04619) find
  69. Liang Wendong, Armin Kekić, Julius von Kügelgen, Simon Buchholz, Michel Besserve, Luigi Gresele, and Bernhard Schölkopf. 2023. Causal Component Analysis. arXiv preprint. ArXiv:2305.17225 [cs, stat]. — [arXiv:2305.17225](https://arxiv.org/abs/2305.17225) find
  70. Thaddäus Wiedemer, Prasanna Mayilvahanan, Matthias Bethge, and Wieland Brendel. 2023. Compositional Generalization from First Principles. arXiv preprint. ArXiv:2307.05596 [cs, stat]. — [arXiv:2307.05596](https://arxiv.org/abs/2307.05596) find
  71. Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D Manning, and Christopher Potts. 2025. AxBench: Steering LLMs? even simple baselines outperform sparse autoencoders. In Forty-second International Conference on Machine Learning. find
  72. Jiaqi Zhang, Chandler Squires, Kristjan Greenewald, Akash Srivastava, Karthikeyan Shanmugam, and Caroline Uhler. 2023. Identifiability Guarantees for Causal Disentanglement from Soft Interventions. arXiv preprint. ArXiv:2307.06250 [cs, math, stat]. — [arXiv:2307.06250](https://arxiv.org/abs/2307.06250) find
  73. Zuheng, Xu, Moksh Jain, Ali Denton, Shawn Whitfield, Aniket Didolkar, Berton Earnshaw, and Jason Hartford. 2024. Automated discovery of pairwise interactions from unstructured data. Preprint, arXiv:2409.07594. — [arXiv:2409.07594](https://arxiv.org/abs/2409.07594) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mueller-2026-isolation-entanglement-when-do-interpretability/