AI Alignment
Part VII: Verification from Inside

Verification from Inside

In brief

Ideas
  • Interpretability reads activations, attributes outputs, and intervenes on internal state; an explanation is a smaller causal model aligned with the network, and passing that test does not single out one (Ch. 43).
  • Superposition stores more features than dimensions; sparse dictionaries built to undo it find readable directions, but no standard measure checks that they are the features the model uses (Ch. 44).
  • Circuits explain narrow tasks in small models and single prompts in large ones, often in a replacement model, not behavior across the inputs an audit must cover (Ch. 45).
  • Directions read and steer concepts such as truth and being tested; inspection finds what is present, not in general what is absent: under cryptographic assumptions, a trainer can plant a backdoor no efficient inspection detects (Ch. 46; Ch. 47).
Evidence
  • Settled: a probe's accuracy does not show that the network uses what it decodes, and patching results depend on the choice of corruption, metric, and ablation (§43.2.4; §43.4.3).
  • Settled: the dictionary found depends on its size, random seed, and architecture, as several groups showed with different methods (§44.6).
  • Settled: many concepts are linearly decodable and steerable in open models, and, as an existence claim, activation monitors can be evaded when optimized against (§46.1; §46.4.1).
  • Contested: whether interpretability can carry safety claims; ambitious, skeptical, and middle positions read the same results differently, and the evidence does not decide (§47.7).
Open
  • Crux: do interpretability tools find planted objectives that black-box methods miss, and keep doing so as models grow? Blind games run independently of the tools' developers would settle it (§47.7).
  • Missing: as of October 2026, a measured miss rate for any interpretability method across planted objectives that include ones trained to evade it, and a shared benchmark of mechanisms planted at current scale (§47.8; §43.8).

This part is about the third of the book's four problems, verification, pursued through the one source of evidence that does not pass through the model's behavior: its internal activations and weights. It teaches the questions interpretability asks and the three kinds of method it uses (reading activations, attributing outputs, and intervening on internal state), with causal abstraction as the standard of a correct explanation. It then teaches superposition as geometry and the sparse dictionaries built to undo it, the circuits and attribution graphs that trace a computation from prompt to output, and the directions that read and steer concepts such as truth, deception, being tested, and character. It closes by asking what inspection can verify: why a correct explanation is not unique, what a proof would add, eliciting latent knowledge as a formal problem, what logic and cryptography rule out, and whether interpretability can carry safety claims. The program is dated throughout: the main line of the community that took the name mechanistic interpretability ran from circuits to superposition to dictionaries between 2020 and 2024, and since 2025, in the book's reading, it has divided into replacement models, readouts in words, task-judged methods, and decompositions of the weights (inference).

The part follows verification from outside (Part VI) because each behavioral method there meets a limit that inside access might pass: a model that behaves differently when tested, reasoning in text that is not the computation, a judge weaker than the judged. It precedes the parts on maintenance because the methods it reports are what monitoring and control would use if they work, and because it ends by asking how far they can be trusted. Two entries of the negative space the book fixed in advance are tested here (Section 1.7): superposition and the non-uniqueness of learned features, which neuroscience's theory of sparse codes partly anticipates, and a principal that reads its agent's activations, which no older theory of delegation contains. The part also shows the loop between the problems at work: a probe or monitor used to train a model changes what it can later detect, so verification from inside is not independent of instillation.

It assumes the residual stream and attention circuits of a transformer, the persona view of a language model, underspecification and phase transitions in training, and the detection statistics of a monitor at a base rate, each recalled with a cross-reference where it is used.

Chapters in this part

  1. 43 Questions and Methods

    Interpretability asks what a network represents, what it computes, and why it produced an output, using access no overseer of a person has ever had. This chapter teaches the three kinds of method (reading activations with probes, attributing outputs with gradients, and intervening on activations), what each can and cannot show, and causal abstraction, the standard that says when an intervention licenses an explanation and why passing it does not make an explanation unique. It reports what neuroscience's methods recovered from a microprocessor whose workings were known, and dates the program as of October 2026, from circuits and dictionaries to the branches the book distinguishes since 2025.

  2. 44 Superposition and Features

    Neurons in trained networks respond to many unrelated things because networks represent more features than they have dimensions. This chapter teaches superposition as geometry (why a space holds far more nearly orthogonal directions than dimensions, why sparsity makes the interference tolerable, and why a nonlinearity is needed), shows in a toy model how the stored features arrange themselves, reports constructions and bounds for computing in superposition, teaches sparse dictionary learning as the tool built to undo it, and reports the evidence that the dictionary it finds is one of many that fit and that dictionaries have lost to simple baselines on several measurement tasks.

  3. 45 Circuits and Attribution Graphs

    A circuit is the part of a network that carries out a behavior. This chapter derives why a transformer's computation splits into paths that can be read from its weights, reports the circuits found by hand in small models and the methods that automate the search, shows that a circuit's score depends on how the rest of the network is removed, teaches the replacement models (transcoders, frozen attention, and error nodes) behind attribution graphs, and reports what those graphs found in one developer's production model and how often they found anything. It closes with circuits that can be checked against ground truth or turned into proofs, and with what circuit work has and has not delivered for verification: so far, circuits for narrow tasks in small models and explanations of single prompts in production models, not of a large model's behavior across the inputs an audit must cover.

  4. 46 Representations, Probes, and Steering

    Many concepts a language model uses can be read from its activations with a probe and changed with a steering vector. This chapter teaches what it means for a concept to be a direction and why the direction that reads a concept is in general not the one that writes it, reports steering and its limits as a control method, recalls the probes built for truth, deception, and evaluation awareness and asks what the forensic standard for validating detectors requires of any activation probe, and reports persona directions, whether models share our concepts, and what models can say about their own internal states.

  5. 47 What Inspection Can and Cannot Verify

    Inspection finds mechanisms; safety arguments need claims that something is absent. This chapter separates the two, shows that explanations passing the standard tests are not unique and that unrestricted alignment maps make the tests vacuous, asks what a proof built from an explanation gives an outside party, states eliciting latent knowledge as a formal problem and heuristic arguments as the route around it, and teaches what logic and cryptography rule out. It closes with the debate over whether interpretability can carry safety claims, which the evidence does not decide.

References for Part VII

218 works cited across this part's chapters.

  1. Abdelnabi, S., Fay, A., Cherubin, G., Salem, A., Fritz, M., and Paverd, A. (2025). Get My Drift? Catching LLM Task Drift with Activation Deltas. 2025 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). Ch. 46
  2. Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., and Kim, B. (2018). Sanity Checks for Saliency Maps. Advances in Neural Information Processing Systems. Ch. 43
  3. Adler, M., and Shavit, N. (2024). On the Complexity of Neural Computation in Superposition. arXiv. preprint Ch. 44
  4. Alain, G., and Bengio, Y. (2016). Understanding intermediate layers using linear classifier probes. arXiv. preprint Ch. 43
  5. Alfonseca, M., Cebrian, M., Fernandez Anta, A., Coviello, L., Abeliuk, A., and Rahwan, I. (2021). Superintelligence Cannot be Contained: Lessons from Computability Theory. Journal of Artificial Intelligence Research. Ch. 47
  6. Ameisen, E., Lindsey, J., Pearce, A., Gurnee, W., Turner, N. L., Chen, B., … Batson, J. (2025). Circuit Tracing: Revealing Computational Graphs in Language Models. Transformer Circuits Thread. non-peer-reviewed Ch. 43 Ch. 45
  7. Amodei, D. (2025). The Urgency of Interpretability. darioamodei.com. non-peer-reviewed Ch. 47
  8. Anthropic (2022). Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  9. Anthropic (2023d). Distributed Representations: Composition & Superposition. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  10. Anthropic (2023e). Interpretability Dreams. Transformer Circuits Thread. non-peer-reviewed Ch. 47
  11. Anthropic (2024m). Simple probes can catch sleeper agents. Anthropic. non-peer-reviewed Ch. 46
  12. Anthropic (2025a). A Toy Model of Interference Weights. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  13. Anthropic (2025aq). System Card: Claude Sonnet 4.5. Anthropic. non-peer-reviewed Ch. 44
  14. Anthropic (2025b). A Toy Model of Mechanistic (Un)Faithfulness. Transformer Circuits Thread. non-peer-reviewed Ch. 45 Ch. 47
  15. Anthropic (2025g). Building and evaluating alignment auditing agents. Anthropic Alignment Science Blog. non-peer-reviewed Ch. 45
  16. Anthropic (2026ai). Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations. Transformer Circuits Thread. non-peer-reviewed Ch. 43 Ch. 46
  17. Anthropic (2026u). Evaluating and mitigating the growing risk of LLM-discovered 0-days. Anthropic. non-peer-reviewed Ch. 46
  18. Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N. (2024). Refusal in Language Models Is Mediated by a Single Direction. Advances in Neural Information Processing Systems. Ch. 46
  19. Arora, S., Li, Y., Liang, Y., Ma, T., and Risteski, A. (2016). Linear Algebraic Structure of Word Senses, with Applications to Polysemy. Transactions of the Association for Computational Linguistics. doi:10.1162/tacl_a_00034. Ch. 44
  20. Arora, A., Wu, Z., Steinhardt, J., and Schwettmann, S. (2026). Language Model Circuits Are Sparse in the Neuron Basis. arXiv. preprint Ch. 45
  21. Bailey, L., Serrano, A., Sheshadri, A., Seleznyov, M., Taylor, J., Jenner, E., … Emmons, S. (2024). Obfuscated Activations Bypass LLM Latent-Space Defenses. arXiv. preprint Ch. 46 Ch. 47
  22. Balesni, M., Korbak, T., and Evans, O. (2024a). Lessons from Studying Two-Hop Latent Reasoning. arXiv. preprint Ch. 45
  23. Barak, B., Goldreich, O., Impagliazzo, R., Rudich, S., Sahai, A., Vadhan, S., and Yang, K. (2001). On the (Im)possibility of Obfuscating Programs. Lecture Notes in Computer Science. Ch. 47
  24. Beckers, S., and Halpern, J. Y. (2019). Abstracting Causal Models. Proceedings of the AAAI Conference on Artificial Intelligence. doi:10.1609/aaai.v33i01.33012678. Ch. 43
  25. Belinkov, Y. (2022). Probing Classifiers: Promises, Shortcomings, and Advances. Computational Linguistics. Ch. 43
  26. Belrose, N., Ostrovsky, I., McKinney, L., Furman, Z., Smith, L., Halawi, D., Biderman, S., and Steinhardt, J. (2023a). Eliciting Latent Predictions from Transformers with the Tuned Lens. arXiv. preprint Ch. 43
  27. Belrose, N., Schneider-Joseph, D., Ravfogel, S., Cotterell, R., Raff, E., and Biderman, S. (2023b). LEACE: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems. Ch. 43 Ch. 46
  28. Bhaskar, A., Wettig, A., Friedman, D., and Chen, D. (2024). Finding Transformer Circuits with Edge Pruning. Advances in Neural Information Processing Systems. Ch. 45
  29. Binder, F. J., Chua, J., Korbak, T., Sleight, H., Hughes, J., Long, R., … Evans, O. (2024). Looking Inward: Language Models Can Learn About Themselves by Introspection. arXiv. preprint Ch. 46
  30. Braun, D., Bushnaq, L., Heimersheim, S., Mendel, J., and Sharkey, L. (2025). Interpretability in Parameter Space: Minimizing Mechanistic Description Length with Attribution-based Parameter Decomposition. arXiv. preprint Ch. 43 Ch. 45
  31. Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., … Olah, C. (2023). Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  32. Bushnaq, L., Heimersheim, S., Goldowsky-Dill, N., Braun, D., Mendel, J., Hänni, K., … Hobbhahn, M. (2024a). The Local Interaction Basis: Identifying Computationally-Relevant and Sparsely Interacting Features in Neural Networks. arXiv. preprint Ch. 45
  33. Bushnaq, L., Braun, D., and Sharkey, L. (2025). Stochastic Parameter Decomposition. arXiv. preprint Ch. 43 Ch. 45
  34. Bussmann, B., Leask, P., and Nanda, N. (2024). BatchTopK Sparse Autoencoders. arXiv. preprint Ch. 44
  35. Casper, S., Ezell, C., Siegmann, C., Kolt, N., Curtis, T. L., Bucknall, B., … Hadfield-Menell, D. (2024a). Black-Box Access is Insufficient for Rigorous AI Audits. The 2024 ACM Conference on Fairness, Accountability, and Transparency. Ch. 43 Ch. 47
  36. Chan, L., Lang, L., and Jenner, E. (2023). Natural Abstractions: Key Claims, Theorems, and Critiques. AI Alignment Forum. non-peer-reviewed Ch. 46
  37. Chanin, D., Wilken-Smith, J., Dulka, T., Bhatnagar, H., Golechha, S., and Bloom, J. (2024). A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders. Neural Information Processing Systems. Ch. 44
  38. Chen, Z., Lau, E., Mendel, J., Wei, S., and Murfet, D. (2023b). Dynamical versus Bayesian Phase Transitions in a Toy Model of Superposition. arXiv. preprint Ch. 44
  39. Chen, R., Arditi, A., Sleight, H., Evans, O., and Lindsey, J. (2025c). Persona Vectors: Monitoring and Controlling Character Traits in Language Models. arXiv. preprint Ch. 46
  40. Chrisman, B., Bushnaq, L., and Sharkey, L. (2025). Identifying Sparsely Active Circuits Through Local Loss Landscape Decomposition. arXiv. preprint Ch. 45
  41. Christiano, P. (2022a). Mechanistic anomaly detection and ELK. Alignment Research Center blog. non-peer-reviewed Ch. 47
  42. Christiano, P., and Xu, M. (2022). ELK prize results. AI Alignment Forum. non-peer-reviewed Ch. 47
  43. Christiano, P., Cotra, A., and Xu, M. (2021). Eliciting latent knowledge: How to tell if your eyes deceive you. Alignment Research Center. non-peer-reviewed Ch. 47
  44. Christiano, P., Neyman, E., and Xu, M. (2022). Formalizing the presumption of independence. arXiv. preprint Ch. 47
  45. Christiano, P., Hilton, J., Lincoln, A., Neyman, E., and Xu, M. (2024). Towards a Law of Iterated Expectations for Heuristic Estimators. arXiv. preprint Ch. 47
  46. Christiano, P., Hilton, J., Lecomte, V., and Xu, M. (2025). Backdoor defense, learnability and obfuscation. Innovations in Theoretical Computer Science Conference (ITCS 2025), LIPIcs vol. 325. doi:10.4230/lipics.itcs.2025.38. Ch. 47
  47. Chughtai, B., Chan, L., and Nanda, N. (2023). A Toy Model of Universality: Reverse Engineering How Networks Learn Group Operations. International Conference on Machine Learning. Ch. 45
  48. Comsa, I. M., and Shanahan, M. (2025). Does It Make Sense to Speak of Introspection in Large Language Models? arXiv. preprint Ch. 46
  49. Conmy, A., Mavor-Parker, A., Lynch, A., Heimersheim, S., and Garriga-Alonso, A. (2023). Towards Automated Circuit Discovery for Mechanistic Interpretability. Advances in Neural Information Processing Systems. Ch. 43 Ch. 45
  50. Cundy, C., and Gleave, A. (2025). Preference Learning with Lie Detectors can Induce Honesty or Evasion. Neural Information Processing Systems. Ch. 46 Ch. 47
  51. Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. (2023). Sparse Autoencoders Find Highly Interpretable Features in Language Models. International Conference on Learning Representations. Ch. 44
  52. Cywiński, B., Ryd, E., Wang, R., Rajamanoharan, S., Nanda, N., Conmy, A., and Marks, S. (2025). Eliciting Secret Knowledge from Language Models. arXiv. preprint Ch. 47
  53. Dalrymple, D. "., Skalse, J., Bengio, Y., Russell, S., Tegmark, M., Seshia, S., … Tenenbaum, J. (2024). Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems. arXiv. preprint Ch. 47
  54. Donoho, D. (2006). Compressed sensing. IEEE Transactions on Information Theory. Ch. 44
  55. Doshi-Velez, F., and Kim, B. (2017). Towards A Rigorous Science of Interpretable Machine Learning. arXiv. preprint Ch. 43
  56. Dumas, C., Wendler, C., Veselovsky, V., Monea, G., and West, R. (2024). Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers. Annual Meeting of the Association for Computational Linguistics. Ch. 46
  57. Dunefsky, J., Chlenski, P., and Nanda, N. (2024). Transcoders Find Interpretable LLM Feature Circuits. Advances in Neural Information Processing Systems. Ch. 45
  58. Elazar, Y., Ravfogel, S., Jacovi, A., and Goldberg, Y. (2021). Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals. Transactions of the Association for Computational Linguistics. doi:10.1162/tacl_a_00359. Ch. 43
  59. Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., … Olah, C. (2021a). A Mathematical Framework for Transformer Circuits. Transformer Circuits Thread. non-peer-reviewed Ch. 43 Ch. 45
  60. Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., … Olah, C. (2022b). Toy Models of Superposition. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  61. Elhage, N., Lasenby, R., and Olah, C. (2023). Privileged Bases in the Transformer Residual Stream. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  62. Engels, J., Michaud, E. J., Liao, I., Gurnee, W., and Tegmark, M. (2024). Not All Language Model Features Are One-Dimensionally Linear. International Conference on Learning Representations. Ch. 44
  63. Farquhar, S., Varma, V., Kenton, Z., Gasteiger, J., Mikulik, V., and Shah, R. (2023). Challenges with unsupervised LLM knowledge discovery. arXiv. preprint Ch. 47
  64. Fiotto-Kaufman, J., Loftus, A. R., Todd, E., Brinkmann, J., Pal, K., Troitskii, D., … Bau, D. (2024). NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals. arXiv. preprint Ch. 43 Ch. 47
  65. Friedman, D., Lampinen, A., Dixon, L., Chen, D., and Ghandeharioun, A. (2024). Interpretability Illusions in the Generalization of Simplified Models. International Conference on Machine Learning. Ch. 47
  66. Ganguli, S., and Sompolinsky, H. (2012). Compressed Sensing, Sparsity, and Dimensionality in Neuronal Information Processing and Data Analysis. Annual Review of Neuroscience. Ch. 44
  67. Gao, L., la Tour, T. D., Tillman, H., Goh, G., Troll, R., Radford, A., … Wu, J. (2024). Scaling and evaluating sparse autoencoders. International Conference on Learning Representations. Ch. 44
  68. Gao, L., Rajaram, A., Coxon, J., Govande, S. V., Baker, B., and Mossing, D. (2025). Weight-sparse transformers have interpretable circuits. arXiv. preprint Ch. 45
  69. Ge, X., Zhu, F., Shu, W., Wang, J., He, Z., and Qiu, X. (2024a). Automatically Identifying Local and Global Circuits with Linear Computation Graphs. arXiv. preprint Ch. 45
  70. Geiger, A., Lu, H., Icard, T., and Potts, C. (2021). Causal Abstractions of Neural Networks. Advances in Neural Information Processing Systems. Ch. 43
  71. Geiger, A., Wu, Z., Potts, C., Icard, T., and Goodman, N. D. (2023). Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations. CLEaR. Ch. 43
  72. Geiger, A., Ibeling, D., Zur, A., Chaudhary, M., Chauhan, S., Huang, J., … Icard, T. (2025). Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability. Journal of Machine Learning Research, 26(83):1−64. Ch. 43 Ch. 47
  73. Ghandeharioun, A., Caciularu, A., Pearce, A., Dixon, L., and Geva, M. (2024). Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models. International Conference on Machine Learning. Ch. 43
  74. Goldowsky-Dill, N., MacLeod, C., Sato, L., and Arora, A. (2023). Localizing Model Behavior with Path Patching. arXiv. preprint Ch. 45
  75. Goldowsky-Dill, N., Chughtai, B., Heimersheim, S., and Hobbhahn, M. (2025). Detecting Strategic Deception Using Linear Probes. arXiv. preprint Ch. 46
  76. Goldwasser, S., Kim, M. P., Vaikuntanathan, V., and Zamir, O. (2022). Planting Undetectable Backdoors in Machine Learning Models : [Extended Abstract]. 2022 IEEE 63rd Annual Symposium on Foundations of Computer Science (FOCS). Ch. 47
  77. Goldwasser, S., Shafer, J., Vafa, N., and Vaikuntanathan, V. (2025). Oblivious Defense in ML Models: Backdoor Removal without Detection. Proceedings of the 57th Annual ACM Symposium on Theory of Computing. doi:10.1145/3717823.3718245. Ch. 47
  78. Gould, R., Ong, E., Ogden, G., and Conmy, A. (2023). Successor Heads: Recurring, Interpretable Attention Heads In The Wild. International Conference on Learning Representations. Ch. 45
  79. Gross, J., Agrawal, R., Kwa, T., Ong, E., Yip, C. H., Gibson, A., Noubir, S., and Chan, L. (2024). Compact Proofs of Model Performance via Mechanistic Interpretability. Advances in Neural Information Processing Systems. doi:10.52202/079017-2523. Ch. 45 Ch. 47
  80. Gupta, R., Arcuschin, I., Kwa, T., and Garriga-Alonso, A. (2024). InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques. arXiv. preprint Ch. 43
  81. Gurnee, W., and Tegmark, M. (2023). Language Models Represent Space and Time. International Conference on Learning Representations. Ch. 46
  82. Gurnee, W., Nanda, N., Pauly, M., Harvey, K., Troitskii, D., and Bertsimas, D. (2023). Finding Neurons in a Haystack: Case Studies with Sparse Probing. Trans. Mach. Learn. Res.. Ch. 44
  83. Gurnee, W., Horsley, T., Guo, Z. C., Kheirkhah, T. R., Sun, Q., Hathaway, W., Nanda, N., and Bertsimas, D. (2024). Universal Neurons in GPT2 Language Models. Trans. Mach. Learn. Res.. Ch. 44
  84. Gurnee, W., Ameisen, E., Kauvar, I., Tarng, J., Pearce, A., Olah, C., and Batson, J. (2025). When Models Manipulate Manifolds: The Geometry of a Counting Task. Transformer Circuits Thread. non-peer-reviewed Ch. 44 Ch. 45
  85. Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., … Lindsey, J. (2026). Verbalizable Representations Form a Global Workspace in Language Models. Transformer Circuits Thread. non-peer-reviewed Ch. 46
  86. Hanna, M., Pezzelle, S., and Belinkov, Y. (2024). Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanisms. arXiv. preprint Ch. 45
  87. Hänni, K., Mendel, J., Vaintrob, D., and Chan, L. (2024). Mathematical Models of Computation in Superposition. arXiv. preprint Ch. 44
  88. Hase, P., Bansal, M., Kim, B., and Ghandeharioun, A. (2023). Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models. Advances in Neural Information Processing Systems. Ch. 43 Ch. 47
  89. He, Z., Shu, W., Ge, X., Chen, L., Wang, J., Zhou, Y., … Qiu, X. (2024c). Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders. arXiv. preprint Ch. 44
  90. He, Z., Achterberg, J., Collins, K., Nejad, K., Akarca, D., Yang, Y., … Lindsay, G. W. (2024d). Multilevel Interpretability Of Artificial Neural Networks: Leveraging Framework And Methods From Neuroscience. arXiv. preprint Ch. 43
  91. Heap, T., Lawson, T., Farnik, L., and Aitchison, L. (2025). Automated Interpretability Metrics Do Not Distinguish Trained and Random Transformers. arXiv. preprint Ch. 44
  92. Heimersheim, S., and Nanda, N. (2024). How to use and interpret activation patching. arXiv. preprint Ch. 43
  93. Hendrycks, D., and Hiscott, L. (2025). The Misguided Quest for Mechanistic AI Interpretability. AI Frontiers. non-peer-reviewed Ch. 47
  94. Henighan, T., Carter, S., Hume, T., Elhage, N., Lasenby, R., Fort, S., Schiefer, N., and Olah, C. (2023). Superposition, Memorization, and Double Descent. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  95. Hernandez, E., Sharma, A. S., Haklay, T., Meng, K., Wattenberg, M., Andreas, J., Belinkov, Y., and Bau, D. (2023). Linearity of Relation Decoding in Transformer Language Models. International Conference on Learning Representations. Ch. 46
  96. Hewitt, J., and Liang, P. (2019). Designing and Interpreting Probes with Control Tasks. Conference on Empirical Methods in Natural Language Processing. doi:10.18653/v1/D19-1275. Ch. 43
  97. Hilton, J. (2024a). A bird's eye view of ARC's research. Alignment Research Center blog. non-peer-reviewed Ch. 47
  98. Hilton, J. (2024b). Formal verification, heuristic explanations and surprise accounting. Alignment Research Center blog. non-peer-reviewed Ch. 47
  99. Hindupur, S. S. R., Lubana, E. S., Fel, T., and Ba, D. (2025). Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry. Neural Information Processing Systems. Ch. 44
  100. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., … Sifre, L. (2022). Training Compute-Optimal Large Language Models. arXiv. preprint Ch. 43
  101. Hoogland, J., Oldenziel, A. G., Murfet, D., and van Wingerden, S. (2023). Towards Developmental Interpretability. AI Alignment Forum. non-peer-reviewed Ch. 43
  102. Huang, V., Choi, D., Johnson, D. D., Schwettmann, S., and Steinhardt, J. (2025b). Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants. arXiv. preprint Ch. 46
  103. Huh, M., Cheung, B., Wang, T., and Isola, P. (2024). The Platonic Representation Hypothesis. International Conference on Machine Learning. Ch. 46
  104. Jacovi, A., and Goldberg, Y. (2020). Towards Faithfully Interpretable NLP Systems: How Should We Define and Evaluate Faithfulness? Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Ch. 43 Ch. 46 Ch. 47
  105. Jenner, E., Kapur, S., Georgiev, V., Allen, C., Emmons, S., and Russell, S. (2024). Evidence of Learned Look-Ahead in a Chess-Playing Neural Network. Advances in Neural Information Processing Systems. doi:10.52202/079017-0987. Ch. 45
  106. Jermyn, A. S., Schiefer, N., and Hubinger, E. (2022). Engineering Monosemanticity in Toy Models. arXiv. preprint Ch. 44
  107. Johnson, W. B., and Lindenstrauss, J. (1984). Extensions of Lipschitz mappings into a Hilbert space. Contemporary Mathematics. Ch. 44
  108. Jonas, E., and Kording, K. P. (2017). Could a Neuroscientist Understand a Microprocessor? PLOS Computational Biology. Ch. 43 Ch. 47
  109. Kamath, H., Ameisen, E., Kauvar, I., Luger, R., Gurnee, W., Pearce, A., … Lindsey, J. (2025). Tracing Attention Computation Through Feature Interactions. Transformer Circuits Thread. non-peer-reviewed Ch. 45
  110. Kantamneni, S., Engels, J., Rajamanoharan, S., Tegmark, M., and Nanda, N. (2025). Are Sparse Autoencoders Useful? A Case Study in Sparse Probing. International Conference on Machine Learning. Ch. 44 Ch. 47
  111. Karvonen, A., Rager, C., Marks, S., and Nanda, N. (2024). Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks. arXiv. preprint Ch. 44
  112. Karvonen, A., Chua, J., Dumas, C., Fraser-Taliente, K., Kantamneni, S., Minder, J., … Marks, S. (2025). Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers. arXiv. preprint Ch. 43 Ch. 46
  113. Kindermans, P.-J., Hooker, S., Adebayo, J., Alber, M., Schütt, K. T., Dähne, S., Erhan, D., and Kim, B. (2017b). The (Un)reliability of saliency methods. Explainable AI. doi:10.1007/978-3-030-28954-6_14. Ch. 43
  114. Kokhlikyan, N., Miglani, V., Alsallakh, B., Martin, M., and Reblitz-Richardson, O. (2021). Investigating sanity checks for saliency maps with image and text classification. arXiv. preprint Ch. 43
  115. Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., and Poeppel, D. (2017). Neuroscience Needs Behavior: Correcting a Reductionist Bias. Neuron. Ch. 43
  116. Kramár, J., Lieberum, T., Shah, R., and Nanda, N. (2024). AtP*: An efficient and scalable method for localizing LLM behaviour to components. arXiv. preprint Ch. 43
  117. Ladner, T., and Althoff, M. (2026). The Misery of Mechanistic Interpretability: A Formal Perspective. arXiv. preprint Ch. 47
  118. Lan, M., Torr, P., Meek, A., Khakzar, A., Krueger, D., and Barez, F. (2024). Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoders. arXiv. preprint Ch. 46
  119. Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W., and Müller, K.-R. (2019). Unmasking Clever Hans predictors and assessing what machines really learn. Nature Communications. Ch. 43
  120. Lazebnik, Y. (2002). Can a biologist fix a radio?—Or, what I learned while studying apoptosis. Cancer Cell. Ch. 43
  121. Leask, P., Bussmann, B., Pearce, M., Bloom, J., Tigges, C., Moubayed, N. A., Sharkey, L., and Nanda, N. (2025). Sparse Autoencoders Do Not Find Canonical Units of Analysis. International Conference on Learning Representations. Ch. 44
  122. Li, K., Hopkins, A. K., Bau, D., Viégas, F., Pfister, H., and Wattenberg, M. (2022a). Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task. International Conference on Learning Representations. Ch. 46
  123. Li, K., Patel, O., Viégas, F., Pfister, H., and Wattenberg, M. (2023a). Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. Advances in Neural Information Processing Systems. doi:10.52202/075280-1797. Ch. 46
  124. Lieberum, T., Rahtz, M., Kramár, J., Nanda, N., Irving, G., Shah, R., and Mikulik, V. (2023). Does Circuit Analysis Interpretability Scale? Evidence from Multiple Choice Capabilities in Chinchilla. arXiv. preprint Ch. 45
  125. Lieberum, T., Rajamanoharan, S., Conmy, A., Smith, L., Sonnerat, N., Varma, V., … Nanda, N. (2024). Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2. Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. Ch. 43 Ch. 44 Ch. 47
  126. Lillicrap, T. P., and Kording, K. P. (2019). What does it mean to understand a neural network? arXiv. preprint Ch. 43 Ch. 47
  127. Lindner, D., Kramár, J., Farquhar, S., Rahtz, M., McGrath, T., and Mikulik, V. (2023). Tracr: Compiled Transformers as a Laboratory for Interpretability. Advances in Neural Information Processing Systems. Ch. 43 Ch. 45
  128. Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits Thread. non-peer-reviewed Ch. 46
  129. Lindsey, J. (2026). Emergent Introspective Awareness in Large Language Models. arXiv. preprint Ch. 46
  130. Lindsey, J., Templeton, A., Marcus, J., Conerly, T., Batson, J., and Olah, C. (2024). Sparse Crosscoders for Cross-Layer Features and Model Diffing. Transformer Circuits Thread. non-peer-reviewed Ch. 44 Ch. 45
  131. Lindsey, J., Gurnee, W., Ameisen, E., Chen, B., Pearce, A., Turner, N. L., … Batson, J. (2025). On the Biology of a Large Language Model. Transformer Circuits Thread. non-peer-reviewed Ch. 43 Ch. 45 Ch. 47
  132. Lipton, Z. C. (2018). The Mythos of Model Interpretability. Queue. Ch. 43
  133. Lu, C., Gallagher, J., Michala, J., Fish, K., and Lindsey, J. (2026). The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models. arXiv. preprint Ch. 46
  134. Makelov, A., Lange, G., and Nanda, N. (2024a). Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching. International Conference on Learning Representations. Ch. 43 Ch. 47
  135. Makelov, A., Lange, G., and Nanda, N. (2024b). Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control. arXiv. preprint Ch. 44
  136. Makhzani, A., and Frey, B. (2013). k-Sparse Autoencoders. arXiv. preprint Ch. 44
  137. Mallen, A., Brumley, M., Kharchenko, J., and Belrose, N. (2023). Eliciting Latent Knowledge from Quirky Language Models. arXiv. preprint Ch. 47
  138. Marin-Llobet, A., and Heimersheim, S. (2026). Individual Parameters in Weight-Sparse Transformers Appear Interpretable. arXiv. preprint Ch. 45
  139. Marks, S., and Tegmark, M. (2023). The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets. arXiv. preprint Ch. 46
  140. Marks, S., Rager, C., Michaud, E. J., Belinkov, Y., Bau, D., and Mueller, A. (2024). Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models. International Conference on Learning Representations. Ch. 45
  141. Marks, S., Treutlein, J., Bricken, T., Lindsey, J., Marcus, J., Mishra-Sharma, S., … Hubinger, E. (2025). Auditing language models for hidden objectives. arXiv. preprint Ch. 43 Ch. 44 Ch. 45 Ch. 47
  142. Marr, D. (1982). Vision: a computational investigation into the human representation and processing of visual information. W.H. Freeman. Ch. 43
  143. McDougall, C., Conmy, A., Rushing, C., McGrath, T., and Nanda, N. (2023). Copy Suppression: Comprehensively Understanding an Attention Head. arXiv. preprint Ch. 45
  144. McGrath, T., Rahtz, M., Kramar, J., Mikulik, V., and Legg, S. (2023). The Hydra Effect: Emergent Self-repair in Language Model Computations. arXiv. preprint Ch. 43
  145. McGuinness, M., Serrano, A., Bailey, L., and Emmons, S. (2025). Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors. arXiv. preprint Ch. 46 Ch. 47
  146. Méloux, M., Maniu, S., Portet, F., and Peyrard, M. (2025). Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? International Conference on Learning Representations. Ch. 43 Ch. 47
  147. Meng, K., Bau, D., Andonian, A., and Belinkov, Y. (2022). Locating and Editing Factual Associations in GPT. Advances in Neural Information Processing Systems. Ch. 43
  148. Merullo, J., Eickhoff, C., and Pavlick, E. (2023a). Circuit Component Reuse Across Tasks in Transformer Language Models. International Conference on Learning Representations. Ch. 45
  149. Merullo, J., Eickhoff, C., and Pavlick, E. (2023b). Language Models Implement Simple Word2Vec-style Vector Arithmetic. North American Chapter of the Association for Computational Linguistics. Ch. 46
  150. Michaud, E. J., Liao, I., Lad, V., Liu, Z., Mudide, A., Loughridge, C., … Tegmark, M. (2024). Opening the AI Black Box: Distilling Machine-Learned Algorithms into Code. Entropy. Ch. 45
  151. Michaud, E. J., Gorton, L., and McGrath, T. (2025). Understanding sparse autoencoder scaling in the presence of feature manifolds. arXiv. preprint Ch. 44
  152. Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. International Conference on Learning Representations. Ch. 46
  153. Miller, J., Chughtai, B., and Saunders, W. (2024). Transformer Circuit Faithfulness Metrics are not Robust. arXiv. preprint Ch. 43 Ch. 45 Ch. 47
  154. Minder, J., Dumas, C., Juang, C., Chugtai, B., and Nanda, N. (2025b). Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuning. Neural Information Processing Systems. doi:10.52202/085713-3551. Ch. 44
  155. Modell, A., Rubin-Delanchy, P., and Whiteley, N. (2025). The Origins of Representation Manifolds in Large Language Models. arXiv. preprint Ch. 46
  156. Nanda, N., Lee, A., and Wattenberg, M. (2023a). Emergent Linear Representations in World Models of Self-Supervised Sequence Models. BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP. Ch. 46
  157. Nanda, N., Chan, L., Lieberum, T., Smith, J., and Steinhardt, J. (2023b). Progress measures for grokking via mechanistic interpretability. International Conference on Learning Representations. Ch. 45
  158. Nanda, N., Engels, J., Conmy, A., Rajamanoharan, S., Chughtai, B., McDougall, C., Kramár, J., and Smith, L. (2025). A Pragmatic Vision for Interpretability. LessWrong. non-peer-reviewed Ch. 43 Ch. 47
  159. National Research Council (2003). The Polygraph and Lie Detection. National Academies Press. Ch. 46
  160. Necula, G. C. (1997). Proof-carrying code. Proceedings of the 24th ACM SIGPLAN-SIGACT symposium on Principles of programming languages - POPL '97. Ch. 47
  161. Nguyen, J., Hoang, K., Attubato, C. L., and Hofstätter, F. (2025). Probing and Steering Evaluation Awareness of Language Models. arXiv. preprint Ch. 46
  162. nostalgebraist (2020). interpreting GPT: the logit lens. LessWrong. non-peer-reviewed Ch. 43
  163. Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., and Carter, S. (2020b). Zoom In: An Introduction to Circuits. Distill. Ch. 43 Ch. 44 Ch. 45
  164. Olshausen, B. A., and Field, D. J. (1996). Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature. Ch. 44
  165. Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., … Olah, C. (2022). In-Context Learning and Induction Heads. Transformer Circuits Thread. non-peer-reviewed Ch. 45
  166. Park, K., Choe, Y. J., and Veitch, V. (2024c). The Linear Representation Hypothesis and the Geometry of Large Language Models. International Conference on Machine Learning. Ch. 46
  167. Paulo, G., and Belrose, N. (2025). Sparse Autoencoders Trained on the Same Data Learn Different Features. arXiv. preprint Ch. 44
  168. Paulo, G., Shabalin, S., and Belrose, N. (2025). Transcoders Beat Sparse Autoencoders for Interpretability. arXiv. preprint Ch. 44 Ch. 45
  169. Pearl, J. (2001). Direct and Indirect Effects. Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence. Ch. 43
  170. Pimentel, T., Valvoda, J., Hall Maudslay, R., Zmigrod, R., Williams, A., and Cotterell, R. (2020). Information-Theoretic Probing for Linguistic Structure. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. doi:10.18653/v1/2020.acl-main.420. Ch. 43
  171. Plunkett, D., Morris, A., Reddy, K., and Morales, J. (2025). Self-Interpretability: LLMs Can Describe Complex Internal Processes that Drive Their Decisions. arXiv. preprint Ch. 46
  172. President's Council of Advisors on Science and Technology (2016). Report to the President: Forensic Science in Criminal Courts: Ensuring Scientific Validity of Feature-Comparison Methods. Executive Office of the President, President's Council of Advisors on Science and Technology. non-peer-reviewed Ch. 46 Ch. 47
  173. Quiroga, R. Q., Reddy, L., Kreiman, G., Koch, C., and Fried, I. (2005). Invariant visual representation by single neurons in the human brain. Nature. Ch. 44
  174. Rajamanoharan, S., Conmy, A., Smith, L., Lieberum, T., Varma, V., Kramár, J., Shah, R., and Nanda, N. (2024a). Improving Dictionary Learning with Gated Sparse Autoencoders. arXiv. preprint Ch. 44
  175. Rajamanoharan, S., Lieberum, T., Sonnerat, N., Conmy, A., Varma, V., Kramár, J., and Nanda, N. (2024b). Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders. arXiv. preprint Ch. 44
  176. Ren, J., Guo, Q., Yan, H., Liu, D., Zhang, Q., Qiu, X., and Lin, D. (2024a). Identifying Semantic Induction Heads to Understand In-Context Learning. Annual Meeting of the Association for Computational Linguistics. Ch. 45
  177. Rentzeperis, I., Calatroni, L., Perrinet, L., and Prandi, D. (2023). Beyond sparse coding in V1. PLoS Comput. Biol.. doi:10.1371/journal.pcbi.1011459. Ch. 44
  178. Rice, H. G. (1953). Classes of recursively enumerable sets and their decision problems. Transactions of the American Mathematical Society. Ch. 47
  179. Rimsky, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., and Turner, A. (2024). Steering Llama 2 via Contrastive Activation Addition. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Ch. 46
  180. Roger, F., Greenblatt, R., Nadeau, M., Shlegeris, B., and Thomas, N. (2023). Benchmarks for Detecting Measurement Tampering. arXiv. preprint Ch. 47
  181. Rubenstein, P. K., Weichwald, S., Bongers, S., Mooij, J. M., Janzing, D., Grosse-Wentrup, M., and Schölkopf, B. (2017). Causal Consistency of Structural Equation Models. arXiv. preprint Ch. 43
  182. Rushing, C., and Nanda, N. (2024). Explorations of Self-Repair in Language Models. International Conference on Machine Learning. Ch. 43
  183. Saphra, N., and Wiegreffe, S. (2024). Mechanistic? arXiv. preprint Ch. 43
  184. Scherlis, A., Sachan, K., Jermyn, A. S., Benton, J., and Shlegeris, B. (2022). Polysemanticity and Capacity in Neural Networks. arXiv. preprint Ch. 44
  185. Segerie, C.-R. (2023). Against Almost Every Theory of Impact of Interpretability. AI Alignment Forum. non-peer-reviewed Ch. 47
  186. Sharkey, L., Chughtai, B., Batson, J., Lindsey, J., Wu, J., Bushnaq, L., … McGrath, T. (2025). Open Problems in Mechanistic Interpretability. arXiv. preprint Ch. 43 Ch. 44 Ch. 47
  187. Sheshadri, A., Ewart, A., Kempf, E., Fronsdal, K., Gupta, I., Schrodi, S., … Wang, R. (2026). AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors. arXiv. preprint Ch. 47
  188. Simonyan, K., Vedaldi, A., and Zisserman, A. (2013). Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. International Conference on Learning Representations. Ch. 43
  189. Smith, L., Rajamanoharan, S., Conmy, A., McDougall, C., Lieberum, T., Kramár, J., Shah, R., and Nanda, N. (2025b). Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research (GDM Mech Interp Team Progress Update #2). AI Alignment Forum. non-peer-reviewed Ch. 43 Ch. 44 Ch. 47
  190. Song, S., Hu, J., and Mahowald, K. (2025a). Language Models Fail to Introspect About Their Knowledge of Language. arXiv. preprint Ch. 46
  191. Song, S., Lederman, H., Hu, J., and Mahowald, K. (2025b). Privileged Self-Access Matters for Introspection in AI. arXiv. preprint Ch. 46
  192. Stolfo, A., Balachandran, V., Yousefi, S., Horvitz, E., and Nushi, B. (2024). Improving Instruction-Following in Language Models through Activation Steering. International Conference on Learning Representations. Ch. 46
  193. Subramani, N., Suresh, N., and Peters, M. (2022). Extracting Latent Steering Vectors from Pretrained Language Models. Findings of the Association for Computational Linguistics: ACL 2022. doi:10.18653/v1/2022.findings-acl.48. Ch. 46
  194. Sundararajan, M., Taly, A., and Yan, Q. (2017). Axiomatic Attribution for Deep Networks. International Conference on Machine Learning. Ch. 43
  195. Sutter, D., Minder, J., Hofmann, T., and Pimentel, T. (2025). The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability? Advances in Neural Information Processing Systems 38. doi:10.52202/085713-5000. Ch. 43 Ch. 47
  196. Syed, A., Rager, C., and Conmy, A. (2023). Attribution Patching Outperforms Automated Circuit Discovery. BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP. Ch. 43 Ch. 45
  197. Templeton, A., Conerly, T., Marcus, J., Lindsey, J., Bricken, T., Chen, B., … Henighan, T. (2024). Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  198. Thompson, K. (1984). Reflections on trusting trust. Communications of the ACM. Ch. 47
  199. Thurnherr, H., and Scheurer, J. (2024). TracrBench: Generating Interpretability Testbeds with Large Language Models. arXiv. preprint Ch. 43
  200. Tigges, C., Hollinsworth, O. J., Geiger, A., and Nanda, N. (2023). Linear Representations of Sentiment in Large Language Models. arXiv. preprint Ch. 46
  201. Tigges, C., Hanna, M., Yu, Q., and Biderman, S. (2024). LLM Circuit Analyses Are Consistent Across Training and Scale. Workshop on Representation Learning for NLP. Ch. 45
  202. Turner, A. M., Thiergart, L., Leech, G., Udell, D., Vazquez, J. J., Mini, U., and MacDiarmid, M. (2023). Steering Language Models With Activation Engineering. arXiv. preprint Ch. 46
  203. Turner, N. L., Wu, J., and Batson, J. (2026). Characterizing interference weights in a tiny language model. Transformer Circuits Thread. non-peer-reviewed Ch. 44
  204. Ulery, B. T., Hicklin, R. A., Buscaglia, J., and Roberts, M. A. (2011). Accuracy and reliability of forensic latent fingerprint decisions. Proceedings of the National Academy of Sciences. Ch. 46 Ch. 47
  205. Vaintrob, D., Mendel, J., and Hänni, K. (2024). Toward A Mathematical Framework for Computation in Superposition. AI Alignment Forum. non-peer-reviewed Ch. 44
  206. Vig, J., Gehrmann, S., Belinkov, Y., Qian, S., Nevo, D., Singer, Y., and Shieber, S. (2020). Investigating Gender Bias in Language Models Using Causal Mediation Analysis. Advances in Neural Information Processing Systems. Ch. 43
  207. Vilas, M. G., Adolfi, F., Poeppel, D., and Roig, G. (2024). Position: An Inner Interpretability Framework for AI Inspired by Lessons from Cognitive Neuroscience. International Conference on Machine Learning. Ch. 43
  208. Voita, E., and Titov, I. (2020). Information-Theoretic Probing with Minimum Description Length. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). doi:10.18653/v1/2020.emnlp-main.14. Ch. 43
  209. Voita, E., Ferrando, J., and Nalmpantis, C. (2023). Neurons in Large Language Models: Dead, N-gram, Positional. Annual Meeting of the Association for Computational Linguistics. Ch. 44
  210. Wang, K., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J. (2023c). Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small. International Conference on Learning Representations. Ch. 43 Ch. 45
  211. Wang, M., la Tour, T. D., Watkins, O., Makelov, A., Chi, R. A., Miserendino, S., … Mossing, D. (2025e). Persona Features Control Emergent Misalignment. arXiv. preprint Ch. 44
  212. Wentworth, J. (2021a). Testing The Natural Abstraction Hypothesis: Project Intro. AI Alignment Forum. non-peer-reviewed Ch. 46 Ch. 47
  213. Wu, G., and Hilton, J. (2025). Estimating the Probabilities of Rare Outputs in Language Models. International Conference on Learning Representations. Ch. 47
  214. Wu, Z., Arora, A., Geiger, A., Wang, Z., Huang, J., Jurafsky, D., Manning, C. D., and Potts, C. (2025b). AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders. International Conference on Machine Learning. Ch. 44 Ch. 46
  215. Yun, Z., Chen, Y., Olshausen, B. A., and LeCun, Y. (2021). Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors. Workshop on Knowledge Extraction and Integration for Deep Learning Architectures; Deep Learning Inside Out. doi:10.18653/V1/2021.DEELIO-1.1. Ch. 44
  216. Zhang, F., and Nanda, N. (2023). Towards Best Practices of Activation Patching in Language Models: Metrics and Methods. International Conference on Learning Representations. Ch. 43 Ch. 45
  217. Zhong, Z., Liu, Z., Tegmark, M., and Andreas, J. (2023). The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks. Advances in Neural Information Processing Systems. Ch. 45
  218. Zou, A., Phan, L., Chen, S., Campbell, J., Guo, P., Ren, R., … Hendrycks, D. (2023a). Representation Engineering: A Top-Down Approach to AI Transparency. arXiv. preprint Ch. 43 Ch. 46