Pith. sign in

REVIEW 5 major objections 5 minor 35 references

A four-stage method extracts sparse computational circuits from language models up to 70 billion parameters, cutting worst-case search from exponential to near-quadratic time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 09:39 UTC pith:FF5HF7KX

load-bearing objection Plausible pipeline, unsupported complexity claim, and a circularity the authors admit: this is exactly the kind of paper that needs a real referee, not a desk reject. the 5 major comments →

arxiv 2601.12879 v2 pith:FF5HF7KX submitted 2026-01-19 cs.LG cs.AIcs.CL

Hierarchical Sparse Circuit Extraction from Billion-Parameter Language Models through Scalable Attribution Graph Decomposition

Mohammed Mudassir Uddin , Shahnawaz Alam , Mohammed Kaif Pasha This is my paper
classification cs.LG cs.AIcs.CL
keywords mechanistic interpretabilitycircuit discoverysparse circuitsattribution graphshierarchical decompositionspectral clusteringgraph neural networkstransformer interpretability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes Hierarchical Attribution Graph Decomposition (HAGD), a four-stage pipeline for extracting sparse computational circuits from transformer language models. It claims to reduce circuit discovery complexity from O(2^n) exhaustive enumeration to O(n^2 log n) by building multi-resolution abstractions of an attribution graph and searching from coarse to fine. The central empirical claim is that circuits of 49 to 347 nodes preserve 82-97% of target behavior on algorithmic tasks and 74-88% on language benchmarks, with up to 91% on modular arithmetic. A sympathetic reader would care because circuit analysis has previously been practical only below one billion parameters; if HAGD holds up, mechanistic interpretability becomes feasible at production scale. The paper is candid that attention-head circuits are omitted and 15-20% of reconstruction variance remains unexplained.

Core claim

The core claim is that circuit discovery can be reformulated as hierarchical search: instead of enumerating all subgraphs, the method first learns cross-layer sparse features, builds a weighted attribution graph from gradient-activation products, coarsens that graph into multiple resolutions via spectral clustering of the normalized Laplacian, and then uses a graph attention network to score which coarse supernodes belong to the circuit, refining only selected branches. The proof of the O(n^2 log n) bound rests on assuming that at each level only O(b) supernodes need consideration. Empirically the paper reports behavioral preservation above 80% on algorithmic and language tasks across model

What carries the argument

The carrying object is the multi-resolution attribution graph hierarchy: spectral coarsening of an edge-weighted graph whose nodes are transcoder features. The hierarchy lets circuit search commit to coarse supernodes before refining, which is what yields the claimed polynomial bound. Causal intervention—ablation for necessity and a circuit-restricted model for sufficiency—verifies the final circuit.

Load-bearing premise

The claim that search costs only O(n^2 log n) depends on the assumption that at each hierarchy level the right circuit can be found by examining only O(b) supernodes and refining those, with no need to consider combinatorial combinations of supernodes.

What would settle it

Run HAGD on a synthetic transformer with a planted circuit whose optimal composition requires nodes from widely separated coarsening branches; if behavioral preservation stays high while the recovered graph omits the planted pathway, the hierarchy search is missing necessary combinations. Alternatively, measure actual wall-clock scaling on graphs of increasing n and check whether it tracks n^2 log n rather than diverging.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the complexity bound holds, circuit extraction at 70B scale becomes routine rather than prohibitive, enabling audits of deployed models.
  • The 82-97% behavioral preservation on arithmetic suggests compact circuits can account for algorithmic behaviors even in large models.
  • Transfer coefficients imply circuits discovered in smaller or different families could seed or constrain searches in larger models, lowering cost further.
  • The stated limitations imply that full mechanistic maps still require attention-circuit modeling and closing the 15-20% reconstruction gap.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The O(b) supernode assumption is doing heavy lifting; a concrete stress test would be to construct a graph where the optimal circuit requires coordinating supernodes across levels that the GNN would not select jointly.
  • If HAGD generalizes, it offers a practical template for 'circuit pretraining': meta-learned search policies on small models could accelerate discovery on large ones, a direction the paper leaves implicit.
  • The reported transfer coefficients align with the possibility of universal algorithmic motifs for formal tasks, but the 18-48% architecture-specific structure warns against assuming transfer is safe for safety-critical conclusions.
  • A testable extension is to apply HAGD to a task with a known ground-truth circuit and compare recovered edges to the known algorithm, directly testing attribution-graph fidelity rather than only behavioral preservation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes HAGD, a four-stage pipeline for extracting sparse computational circuits from large transformers: cross-layer transcoders for monosemantic features, spectral coarsening of attribution graphs, a GNN-guided hierarchical search, and causal intervention. It claims reduction of circuit search from O(2^n) to O(n^2 log n), and reports behavioral-preservation results up to 91% on modular arithmetic with circuits of 49–347 nodes across GPT-2, Pythia, and Llama models, plus cross-architecture transfer coefficients. The abstract and discussion acknowledge limitations including omitted attention-head circuits, 15–20% unexplained reconstruction variance, validation circularity, and interpretability limits of large circuits.

Significance. If the complexity and faithfulness claims were valid, HAGD would be a substantial advance in scaling mechanistic interpretability to production-scale models. The paper also contains candid limitations sections and an interesting attempt at cross-architecture transfer. However, the central complexity proof is internally invalid, and the evaluation uses the same objective for selection and validation; these issues make the main contribution unsupported as written. The empirical comparisons are too limited to compensate: no code, few baselines, no statistical details, and the ablation table contradicts the claimed efficiency benefit.

major comments (5)
  1. [Section III.D, Theorem 1] Theorem 1's proof is invalid. It asserts 'At each hierarchy level, circuit search considers O(b) supernodes for inclusion,' but deciding which O(b) supernodes form the circuit is exactly the combinatorial core of circuit discovery; no argument shows that spectral clustering or the GNN avoids exponential search over supernode subsets. The algebra is also wrong: O(b log_b n) = O(log n) for constant b, not O(n). The statement 'O(n/logb · logb) = O(n)' conflates b with n. Consequently the claimed worst-case O(n^2 log n) complexity, and the scalability claims in Fig. 4 and Table VI, rest on an unproven assumption. The 'edge verification' O(n^2) is a post-selection verification cost, not a discovery cost.
  2. [Section II.C/III.F, Eqs. (2)/(10)] The central validation metric is the same quantity used as the search objective. Eq. (2) defines the discovered circuit as argmin |C| subject to phi(C) >= theta, where phi is 'behavioral preservation.' Eq. (10) then reports phi(C)=Acc(M_C)/Acc(M_full) as sufficiency evidence. High preservation is therefore expected by construction and does not independently establish that the extracted circuit is faithful. Tables II and III's Pres. columns inherit this circularity. The paper needs an out-of-sample or ground-truth-based evaluation—e.g., comparing recovered circuits to known circuits on synthetic tasks or testing on behaviors not used during search.
  3. [Section III.E / Table IV] The GNN meta-model requires ground-truth circuit labels from exhaustive search on small models or manual annotation. This training/labeling cost is not included in the claimed end-to-end complexity and may dominate the total. Moreover, cross-architecture transfer (Table IV) relies on aligning features across models using activation correlations (Eq. 11), but no validation shows that such alignment yields meaningful edge correspondences; the transfer coefficients may reflect trivial overlap due to similar token embeddings. These issues undercut both the complexity and transfer claims.
  4. [Section V.D, Table V] The ablation data are internally inconsistent with the text. Table V reports time 0.3× for 'No Hierarchy' relative to Full Framework at 1.0×, i.e., removing the hierarchy is 70% faster, yet §V.D states that removing hierarchical decomposition 'eliminates the computational efficiency advantage.' Similarly, 'No Causal Valid.' has time 1.4×, which would mean removing causal validation increases runtime. This contradicts the paper's complexity narrative and needs correction; it also does not demonstrate that the hierarchical search finds optimal or near-optimal circuits, only that its quality is comparable on the single GPT-2 Small case.
  5. [Section V.A / Tables II–IV] No confidence intervals, number of seeds, or statistical tests are reported for the preservation and transfer numbers, aside from a single ±2.3% figure in the abstract. Given the small circuits and high variance in interpretability work, the claimed differences between methods (e.g., 0.87 vs 0.91) may not be significant. The exhaustive baseline is run only at GPT-2 Small, so the claim that hierarchical search recovers near-optimal circuits at scale is untested. This is not a fatal issue by itself, but combined with the above it weakens the empirical support.
minor comments (5)
  1. [Abstract and throughout] Exponents are mangled: 'O(2n)' should be 'O(2^n)', and 'O(n 2 logn)' should be 'O(n^2 log n)' in the abstract and Section I. Similar formatting errors appear in the proof.
  2. [Section III.C, Eq. (5)] The attribution definition is dimensionally inconsistent: partial derivative ∂f_j/∂f_i is meaningful only for scalar features, but f_j is described as a feature vector. The index structure and aggregation over positions/layers need to be specified.
  3. [Section III.D proof notation] The expression 'O(n/logb · logb)' is undefined and should be 'O(b log_b n)'. For constant b this is O(log n), not O(n); the authors need to define whether b is constant or grows with n.
  4. [Table V] The 'Time' column uses relative multipliers without defining the baseline unit or whether lower is better. The values 0.3× and 1.4× appear to contradict the accompanying text and should be clarified.
  5. [Reproducibility] No code release, hardware details, GNN hyperparameters, number of training seeds, or preprocessing details for the RedPajama corpus are provided. This makes the empirical claims hard to verify.

Circularity Check

1 steps flagged

Reported behavioral preservation is the same objective used to select circuits; the O(n^2 log n) claim is unsupported but not a circular identity.

specific steps
  1. self definitional [Section II.C Eq. (2); Section III.F Eq. (10); Section VI.C Limitations ('Validation Circularity')]
    "C∗ = arg min_{C⊆G} |C| subject to ϕ(C)≥θ (2) ... ϕ(C) = Acc(MC)/Acc(Mfull) (10) ... Validation Circularity.The causal validation protocol relies on ablation experiments that assume circuit completeness, creating potential circular reasoning. Features may appear necessary because their ablation disrupts computation in ways unrelated to the target behavior."

    Equation (2) defines the discovered circuit as the minimal subgraph satisfying behavioral preservation ϕ(C)≥θ, and Eq. (10) defines behavioral preservation as the accuracy ratio of the circuit-restricted model to the full model. The paper then reports ϕ(C) (e.g., 0.91 on modular arithmetic) as sufficiency evidence and treats ϕ(C)>0.9 as sufficient. The reported preservation is therefore the same optimization/acceptance criterion used to select, prune, and validate the circuit; it is not an independent test. The paper's own Limitation section explicitly concedes this 'circular reasoning.'

full rationale

The clearest circular reduction is in the validation protocol: Eq. (2)'s objective and Eq. (10)'s reported preservation are the same quantity ϕ, and the limitations section explicitly acknowledges 'Validation Circularity.' This makes the headline preservation numbers partly self-confirming, so the score is elevated to 6. The Theorem 1 O(n^2 log n) bound is not circular in the equivalence sense: the proof assumes that search considers O(b) supernodes per level and then contains an algebraic step O(b log_b n)=O(n) that is false for constant b, with no argument ruling out combinatorial search over supernode combinations. That is a correctness/rigor problem rather than an identity between input and output. Similarly, the GNN's need for ground-truth labels from exhaustive search is an acknowledged cost/scaling gap, but since those labels are produced on small models and then applied to new models, it is not a strict 'fitted parameter renamed as prediction' circularity. No load-bearing self-citations are present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central complexity claim rests on unproven hierarchical-optimality and GNN-generalization assumptions. The reported preservation numbers depend on a threshold theta and a behavioral-preservation metric that also drive circuit selection. Many key hyperparameters are unspecified, and the circuit-restricted submodel is not concretely defined.

free parameters (6)
  • sparsity k = 32-128
    TopK active features per hidden state, set per model scale without a reported selection criterion (Section III.B).
  • regularization weights lambda_1, lambda_2 = unspecified
    Balance reconstruction, cross-layer prediction, and L1 sparsity in Eq (4); values never reported.
  • dictionary expansion factor m/d = 8
    Chosen for all models in Section IV.A; no sweep or justification.
  • branching factor b = unspecified
    Used in the complexity proof of Theorem 1; the value is never specified, making the O(n^2 log n) bound non-computable.
  • behavioral preservation threshold theta = unspecified
    Minimum acceptable preservation in Eq (2); not reported, so the optimization problem is under-specified.
  • GNN hyperparameters = unspecified
    Message-passing layers, attention dimensions, and training details for the GNN meta-model are not given.
axioms (5)
  • domain assumption Transformer computation decomposes into sparse, causally meaningful feature graphs.
    Section II.A; the entire circuit abstraction assumes features and edges capture the model's computation, which is the defining hypothesis of mechanistic interpretability.
  • ad hoc to paper Normalized-Laplacian spectral clustering of the attribution graph preserves circuit-relevant structure.
    Section III.D; no theorem or experiment shows that optimal circuits survive the coarsening steps.
  • ad hoc to paper Selecting O(b) supernodes per level and refining them yields the optimal circuit.
    Section III.D, Theorem 1 proof; this is the load-bearing, unproven assumption behind the O(n^2 log n) complexity claim.
  • ad hoc to paper The GNN meta-model, trained on small-model exhaustive labels, generalizes to larger and cross-family models.
    Section III.E; no evidence is provided that label quality or GNN predictions transfer across model scales and families.
  • domain assumption Ablation effects and behavioral-preservation ratios measure causal necessity and sufficiency.
    Section III.F; standard in the field but known to be confounded by distribution shift and circularity, as acknowledged in Section VI.C.
invented entities (2)
  • Multi-resolution supernode hierarchy (coarsened attribution graph) no independent evidence
    purpose: Reduces the circuit search space from exponential to polynomial by grouping related features into macro-nodes
    An internal abstraction with no falsifiable handle outside the pipeline; its validity is assumed, not demonstrated.
  • Circuit-restricted submodel M_C no independent evidence
    purpose: Used to compute behavioral preservation in Eq (10); represents the model operating only on the discovered circuit
    Never defined constructively; unclear how a submodel that omits attention can reproduce attention-dependent behaviors.

pith-pipeline@v1.3.0-alltime-deepseek · 10422 in / 9949 out tokens · 95778 ms · 2026-08-03T09:39:15.544451+00:00 · methodology

0 comments
read the original abstract

Extracting sparse circuits from billion-parameter transformers is constrained by $O(2^n)$ search cost and pervasive feature reuse across co-active pathways. Hierarchical Attribution Graph Decomposition (HAGD) addresses this through four stages: cross-layer transcoder training, spectral coarsening of attribution graphs, graph-neural-network (GNN)-guided hierarchical traversal, and causal intervention verification, reducing worst-case complexity to $O(n^2 \log n)$. Per-layer transcoders trained on the RedPajama corpus yield monosemantic dictionaries; gradient-activation products form weighted attribution graphs; normalized-Laplacian spectral clustering builds multi-resolution hierarchies; an attention-based GNN assigns circuit-membership scores at successive coarsening stages. Evaluation spans GPT-2 (117M-774M), Pythia (1.4B-6.9B), and Llama (7B-70B) across modular arithmetic, parity computation, integer sorting, coreference resolution (WinoGrande), commonsense reasoning (HellaSwag), and factual recall. Behavioral preservation reaches 91\% ($\pm$2.3\%) on modular arithmetic with 49-347-node circuits, while ACDC exhausts memory beyond 1.4B parameters. Cross-architecture transfer coefficients span 0.38-0.82, with within-family pairs (Llama-7B $\to$ Llama-70B) attaining 0.82. Limitations include omitted attention-head circuits, 15-20\% unexplained reconstruction variance, ablation-based validation circularity, and uncertain interpretability of circuits exceeding several hundred nodes.

Figures

Figures reproduced from arXiv: 2601.12879 by Mohammed Kaif Pasha, Mohammed Mudassir Uddin, Shahnawaz Alam.

Figure 1
Figure 1. Figure 1: Hierarchical circuit extraction pipeline from language [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cross-architecture circuit transfer coefficients averaged [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Behavioral preservation versus circuit size across task [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Complexity comparison showing polynomial scaling of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 10 linked inside Pith

  1. [1]

    On the biology of a large language model,

    Anthropic, “On the biology of a large language model,”Trans- former Circuits Thread, 2025. [Online]. Available: https://transformer- circuits.pub/2025/attribution-graphs/methods.html

    2025

  2. [2]

    Neuron-level circuits: Pruning MLPs to interpret their weights,

    A. Bhaskar, D. Friedman, and N. Nanda, “Neuron-level circuits: Pruning MLPs to interpret their weights,”Transluce Research, 2025. [Online]. Available: https://transluce.org/neuron-circuits

    2025

  3. [3]

    Towards monosemanticity: Decomposing language models with dictionary learning,

    T. Brickenet al., “Towards monosemanticity: Decomposing language models with dictionary learning,”Transformer Circuits Thread, 2023

    2023

  4. [4]

    Towards automated circuit discovery for mechanistic interpretability,

    A. Conmy, A. N. Mavor-Parker, A. Lynch, S. Heimersheim, and A. Garriga-Alonso, “Towards automated circuit discovery for mechanistic interpretability,” inAdvances in Neural Information Processing Systems, vol. 36, 2023

    2023

  5. [5]

    Sparse autoencoders find highly interpretable features in language models,

    H. Cunningham, A. Ewart, L. Riggs, R. Huben, and L. Sharkey, “Sparse autoencoders find highly interpretable features in language models,” in Proc. Int. Conf. Learning Representations (ICLR), 2024

    2024

  6. [6]

    A mathematical framework for transformer circuits,

    N. Elhageet al., “A mathematical framework for transformer circuits,” Transformer Circuits Thread, 2022

    2022

  7. [7]

    Scaling and evaluating sparse autoencoders,

    L. Gaoet al., “Scaling and evaluating sparse autoencoders,”arXiv preprint arXiv:2406.04093, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Weight-sparse transformers enable circuit- level interpretability,

    L. Gao, S. Cheng, and J. Wu, “Weight-sparse transformers enable circuit- level interpretability,”OpenAI Research, 2025

    2025

  9. [9]

    How does GPT-2 compute greater- than? Interpreting mathematical abilities in a pre-trained language model,

    M. Hanna, O. Liu, and A. Variengien, “How does GPT-2 compute greater- than? Interpreting mathematical abilities in a pre-trained language model,” inAdvances in Neural Information Processing Systems, vol. 36, 2023

    2023

  10. [10]

    Sparse feature circuits: Discovering and editing interpretable causal graphs in language models,

    S. Marks, C. Rager, E. J. Michaud, Y . Belinkov, D. Bau, and A. Mueller, “Sparse feature circuits: Discovering and editing interpretable causal graphs in language models,”arXiv preprint arXiv:2403.19647, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Locating and editing factual associations in GPT,

    K. Meng, D. Bau, A. Andonian, and Y . Belinkov, “Locating and editing factual associations in GPT,” inAdvances in Neural Information Process- ing Systems, vol. 35, 2022

    2022

  12. [12]

    Opening the AI black box: Program synthesis via mechanistic interpretability,

    E. J. Michaudet al., “Opening the AI black box: Program synthesis via mechanistic interpretability,”arXiv preprint arXiv:2402.05110, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Progress measures for grokking via mechanistic interpretability,

    N. Nanda, L. Chan, T. Liberum, J. Smith, and J. Steinhardt, “Progress measures for grokking via mechanistic interpretability,” inProc. Int. Conf. Learning Representations (ICLR), 2023

    2023

  14. [14]

    Zoom in: An introduction to circuits,

    C. Olah, N. Cammarata, L. Schubert, G. Goh, M. Petrov, and S. Carter, “Zoom in: An introduction to circuits,”Distill, 2020. [Online]. Available: https://doi.org/10.23915/distill.00024

    doi:10.23915/distill.00024 2020

  15. [15]

    In-context learning and induction heads,

    C. Olssonet al., “In-context learning and induction heads,”Transformer Circuits Thread, 2022

    2022

  16. [16]

    Interpretability in weight-sparse language models,

    OpenAI, “Interpretability in weight-sparse language models,”OpenAI Research, 2024. [Online]. Available: https://openai.com/index/understanding-neural-networks-through-sparse- circuits/

    2024

  17. [17]

    WinoGrande: An adversarial Winograd schema challenge at scale,

    K. Sakaguchi, R. Le Bras, C. Bhagavatula, and Y . Choi, “WinoGrande: An adversarial Winograd schema challenge at scale,” inProc. AAAI Conf. Artificial Intelligence, vol. 34, no. 05, 2020, pp. 8732–8740

    2020

  18. [18]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” inProc. Int. Conf. Machine Learning (ICML), 2017, pp. 3319– 3328

    2017

  19. [19]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” inProc. Int. Conf. Learning Representations (ICLR), 2018

    2018

  20. [20]

    A tutorial on spectral clustering,

    U. von Luxburg, “A tutorial on spectral clustering,”Statistics and Com- puting, vol. 17, no. 4, pp. 395–416, 2007

    2007

  21. [21]

    Interpretability in the wild: A circuit for indirect object identification in GPT-2 small,

    K. Wang, A. Variengien, A. Conmy, B. Shlegeris, and J. Steinhardt, “Interpretability in the wild: A circuit for indirect object identification in GPT-2 small,” inProc. Int. Conf. Learning Representations (ICLR), 2023

    2023

  22. [22]

    HellaSwag: Can a machine really finish your sentence?

    R. Zellers, A. Holtzman, Y . Bisk, A. Farhadi, and Y . Choi, “HellaSwag: Can a machine really finish your sentence?” inProc. Annual Meeting of the Association for Computational Linguistics (ACL), 2019, pp. 4791– 4800

    2019

  23. [23]

    Defining and quantifying the emergence of sparse concepts in DNNs,

    Q. Ren, J. Li, H. Zheng, and Q. Zhang, “Defining and quantifying the emergence of sparse concepts in DNNs,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2023, pp. 20280– 20289

    2023

  24. [24]

    Explaining generalization power of a DNN using interactive concepts,

    Q. Ren and Q. Zhang, “Explaining generalization power of a DNN using interactive concepts,”arXiv preprint arXiv:2302.13091, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Where we have arrived in proving the emergence of sparse symbolic concepts in AI models,

    Q. Ren, H. Chen, and Q. Zhang, “Where we have arrived in proving the emergence of sparse symbolic concepts in AI models,”arXiv preprint arXiv:2305.01939, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Discovering transformer circuits via a hybrid attribution and pruning framework,

    Y . Zhang, Z. Li, and M. Chen, “Discovering transformer circuits via a hybrid attribution and pruning framework,”arXiv preprint arXiv:2510.03282, 2025

    arXiv 2025

  27. [27]

    Finding transformer circuits with edge pruning,

    A. Syed, C. Rager, and D. Bau, “Finding transformer circuits with edge pruning,”arXiv preprint arXiv:2406.16778, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    The local interaction basis: Identifying computationally-relevant and sparsely interacting features in neural networks,

    N. Goldowsky-Dill, C. MacLeod, L. Kreiman, and J. Steinhardt, “The local interaction basis: Identifying computationally-relevant and sparsely interacting features in neural networks,”arXiv preprint arXiv:2405.10928, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Functional faithfulness in the wild: Circuit discovery with differentiable computation graph pruning,

    J. Huang, A. Geiger, K. D’Oosterlinck, Z. Wu, and C. Potts, “Functional faithfulness in the wild: Circuit discovery with differentiable computation graph pruning,”arXiv preprint arXiv:2407.03779, 2024

    arXiv 2024

  30. [30]

    Automatically identifying local and global circuits with linear computation graphs,

    A. Makelov, G. Lange, and N. Nanda, “Automatically identifying local and global circuits with linear computation graphs,”arXiv preprint arXiv:2405.13868, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Jacobian sparse autoen- coders: Sparsify computations, not just activations,

    L. Riggs, E. J. Michaud, and A. Conmy, “Jacobian sparse autoen- coders: Sparsify computations, not just activations,”arXiv preprint arXiv:2502.18147, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Weight-sparse transformers have interpretable circuits,

    L. Gao, A. Rajaram, J. Coxon, S. V . Govande, B. Baker, and D. Mossing, “Weight-sparse transformers have interpretable circuits,”arXiv preprint arXiv:2511.13653, 2025

    arXiv 2025

  33. [33]

    Language models can explain neurons in language models,

    S. Bills, N. Cammarata, D. Mossing, H. Tillman, L. Gao, G. Goh, I. Sutskever, J. Leike, J. Wu, and W. Saunders, “Language models can explain neurons in language models,”OpenAI Research, 2023

    2023

  34. [34]

    Scaling monosemanticity: Extracting interpretable features from Claude 3 Sonnet,

    A. Templetonet al., “Scaling monosemanticity: Extracting interpretable features from Claude 3 Sonnet,”Transformer Circuits Thread, 2024

    2024

  35. [35]

    Does circuit analysis interpretability scale? Evi- dence from multiple choice capabilities in Chinchilla,

    T. Lieberum, M. Rahtz, J. Kram ´ar, N. Nanda, G. Irving, R. Shah, and V . Mikulik, “Does circuit analysis interpretability scale? Evi- dence from multiple choice capabilities in Chinchilla,”arXiv preprint arXiv:2307.09458, 2023

    Pith/arXiv arXiv 2023