Pith. sign in

REVIEW 4 major objections 4 minor 174 references

Targeted parameter decomposition recovers the exact weight circuits behind chosen prompts, then shows those circuits can be ablated or rewired in a 12-block transformer with negligible side effects.

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

2026-08-02 10:35 UTC pith:LHEZFEMK

load-bearing objection tPD is a genuinely useful extension of parameter decomposition, but the "mechanistically faithful" claim is overbroad—their own Appendix E shows the catch-all residual can hide redundant target-relevant mechanisms. the 4 major comments →

arxiv 2607.13047 v1 pith:LHEZFEMK submitted 2026-06-19 cs.LG

Targeted Recovery of Weight-Space Mechanisms From Neural Networks

Antoine Vigouroux , Lee Sharkey This is my paper
classification cs.LG
keywords mechanistic interpretabilityparameter decompositiontargeted decompositionrank-1 subcomponentscircuit editingknowledge unlearningtransformer interpretabilitycatch-all residual
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 introduces targeted parameter decomposition (tPD), a way to reverse-engineer only the mechanisms in a neural network that process a chosen set of inputs, rather than decomposing the whole model. It does this by training the decomposition on two parallel data streams: a sparse set of rank-1 subcomponents must reconstruct the target outputs, while a full-rank catch-all component Δ absorbs everything else and is adversarially disabled on target batches. On a 4-block Pile-trained transformer, tPD extracts a CSS-only submodel using about 7% of the FLOPs of the full decomposition, with comparable reconstruction. On a 12-block transformer, ablating prompt-specific subcomponents erases the 'np'/'pd' completions, and swapping their output directions flips the completions as predicted, with per-token KL divergence below 10^-2 on unrelated Pile tokens. The authors claim these recovered components are the mechanisms the original model actually uses, not just an alternative implementation.

Core claim

The central claim is that a transformer's behaviour on a specific input — a prompt, a set of prompts, or a whole language subtask — can be fully accounted for by a small set of rank-1 weight subcomponents U_i V_i^T, and that these subcomponents can be recovered cheaply and then surgically modified. The key device is the catch-all residual Δ = W − Σ U_i V_i^T: it is forced to be inactive on target inputs via adversarial ablation, so any mechanism that fires on the target must live in the visible subcomponents. The paper validates this on toy models (recovering exactly the three mechanisms for features 5, 10, 15), on reproducibility and nesting checks in a 4-block transformer, and on a 12-bloc

What carries the argument

The load-bearing structure is the per-matrix decomposition W ≈ Σ_i m_i U_i V_i^T + m_Δ (W − Σ_i U_i V_i^T), where U_i and V_i are the output and input directions of rank-1 subcomponents, m_i are masks set by a causal-importance network, and Δ is a full-rank catch-all. On target batches the masks are chosen adversarially to maximize reconstruction loss, with Δ ablated between 0% and 100%, so target outputs cannot rely on Δ; on non-target batches Δ is always enabled and subcomponent masks are sampled to minimise their interference. An importance-minimality loss on both streams drives the subcomponents to be as few and as sparse as possible. This two-stream, adversarial design is what lets tPD

Load-bearing premise

The load-bearing assumption is that no target-relevant mechanism is hidden inside the catch-all Δ; if a redundant mechanism fires on both target and non-target inputs, the training objective can relegate it to Δ, and the recovered circuit silently misses part of the computation — something the paper's own Appendix E reports preliminary evidence for in the CSS-only case.

What would settle it

Find a subcomponent whose ablation in the decomposed circuit causes a large KL increase but whose ablation in the full model causes almost none (the paper's Appendix E test). If such a component exists in the numpy/pandas editing case — where the paper currently reports none — the claim that the relevant memory was fully captured and erased would fail. More directly: train a toy transformer with two redundant copying circuits (two attention layers, each dropped independently with probability 0.4 during training), run tPD on the copying task, and check whether one of the two circuits is missing

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

If this is right

  • tPD recovers a CSS-only submodel of a 4-block transformer using ≈7% of the FLOPs of the full-data decomposition, with mean adversarial KL 0.45 versus 0.48 for the full run, suggesting targeted explanation can match or beat full decomposition on the slice of behaviour it covers.
  • In a 12-block transformer, ablating the prompt-specific subcomponents suppresses the 'import numpy as' and 'import pandas as' completions while leaving >100k tokens of Pile data essentially untouched (per-token KL < 10^-2).
  • Swapping the U output directions of two pairs of prompt-specific rank-1 subcomponents flips the completions in the predicted way — 'numpy' now completes with 'pd' and 'pandas' with 'np' — including inside natural code contexts.
  • Decompositions are reproducible across random seeds and coherent across nested target sets: 69 of 77 subcomponents found for {np, pd} match subcomponents found for {np} or {pd} alone.
  • Because the method is unsupervised (inputs are specified, not tasks or metrics), it can be applied to any targeted question about a model — localising, erasing, or editing specific behaviours without presupposing where or how knowledge is stored.

Where Pith is reading between the lines

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

  • If the 12-block swapping result generalises beyond token-specific memorised sequences, tPD offers a path to knowledge editing built directly on weight-space mechanisms: instead of imposing an edit on activations or assuming a key-value store, one would reverse-engineer the rank-1 connections and modify them.
  • The catch-all Δ also points to a probe for redundancy: by comparing how much a subcomponent's ablation hurts the full model versus the extracted circuit (the paper's Appendix E test), one can detect backup or self-repair mechanisms that single-component ablations mask. A natural next experiment is to apply that test at scale and ask whether 'erased' facts are truly gone or merely relocated to Δ.
  • A testable extension we draw beyond the paper: run tPD on the same target with different non-target distributions. If the recovered subcomponents and their ablation effects shift dramatically, then the claimed 'mechanisms' are partly artifacts of the background stream; if they are stable, the target mechanisms are robustly identified.

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

4 major / 4 minor

Summary. The paper introduces Targeted Parameter Decomposition (tPD), a variant of parameter decomposition that recovers only the weight-space mechanisms active on a user-chosen target dataset. The method adds a high-rank catch-all component Δ that is fully enabled on non-target data and adversarially ablated on target data, forcing target behavior to be explained by a sparse set of rank-1 subcomponents. The authors validate tPD on a toy model (TMCC), a 4-block Pile-trained transformer (numpy/pandas prompts and CSS code), and a 12-block Pile transformer (numpy/pandas editing). They report compute savings (≈7% FLOPs for the CSS-only decomposition), seed-consistency, nested-set consistency, and a swap experiment in which swapping U directions of two selected subcomponents flips the model's completions from 'np' to 'pd' and vice versa with negligible side effects. The central claim is that tPD recovers 'reproducible, mechanistically faithful circuits' that support surgical ablation and rewiring.

Significance. If the claims hold, tPD would be a practical tool for narrow-scope mechanistic interpretability, unlearning, and editing at scales where full-data PD is infeasible. The toy-model ground truth is clean, the code is provided, and the swap experiment is a genuine out-of-sample test that the objective does not explicitly optimize. The compute reduction (≈7% FLOPs on the CSS case) is meaningful, and the seed-consistency checks with a null distribution are a good practice. However, the central 'faithful circuits' claim is not fully supported: the paper's own Appendix E reports preliminary evidence that the CSS-only decomposition misses redundant mechanisms hidden in Δ. This limits the strength of the unlearning/editing conclusions, which require completeness, not just causal sufficiency.

major comments (4)
  1. [§2, Eq. (1); Appendix E] The abstract and Section 1 claim that tPD 'recovers reproducible, mechanistically faithful circuits', but the method does not guarantee that target-relevant mechanisms are not absorbed into the catch-all Δ. The objective adversarially ablates Δ on target batches, but nothing prevents a mechanism that is genuinely used on target inputs from being represented in Δ, as long as the subcomponents alone can reconstruct target outputs. The importance-minimality loss actively encourages pushing redundant mechanisms into Δ, since Δ is not counted. Appendix E explicitly states 'there is preliminary evidence that the CSS-only decomposition misses some redundant mechanisms hidden in the Δ components' (Fig. A3). This directly undercuts the faithfulness claim and the unlearning motivation: ablating recovered subcomponents leaves Δ mechanisms intact. Please either weaken the claims to 'minimal sufficie
  2. [§3.4, Fig. 4] The swap experiment is strong evidence of causal sufficiency for the two selected subcomponent pairs, but it does not establish completeness. With only two pairs, chosen 'a priori' by the authors, and no redundancy check on the 12-block model, the experiment cannot rule out that additional mechanisms in Δ contribute to the same np/pd completions. The text says 'the off-diagonal subcomponents are all highly specific' and 'Presumably, these subcomponents represent where these associations are stored', but this is a sufficiency statement, not a faithfulness/completeness statement. The claim 'surgically ablate and rewire memorized sequences' should be qualified—e.g., 'a sufficient circuit for the target completions'—unless the authors demonstrate that no target-relevant mechanism remains in Δ for these sequences.
  3. [§3.2, Fig. 1E] The nested-set consistency check reports that 69 of 77 joint-decomposition subcomponents have a 'strong match' in the single-prompt decompositions, and that 32 are 'seemingly used by both prompts'. Unlike the seed-consistency check in Fig. 1D, no null distribution, matching threshold, or quantitative criterion is provided for Fig. 1E. Without this, the reader cannot assess whether 69/77 is significantly above chance, especially given the component-merging phenomenon described in Appendix C. Please report the matching metric and threshold, and for the 'shared by both' category show the actual overlap distribution or a null baseline.
  4. [§3.3, Fig. 2A] The CSS-only submodel is presented as 'behaviorally similar to the original on CSS data (KL-div. ≈ 0.6)', but the figure shows that per-token KL can be substantially higher on some CSS tokens, and the mean is only an order of magnitude below the collapsed-language values. The claim 'match and exceed its reconstruction accuracy' for the CSS decomposition is based on a single scalar (mean adversarial KL-divergence 0.45), which may hide important failure modes. Please report the distribution (e.g., percentiles) and clarify whether the CSS-only model's outputs on CSS are actually close to the original in the tokens that matter for the claimed syntactic mechanisms.
minor comments (4)
  1. [Abstract and §1] Consider replacing 'faithfully reflect' and 'mechanistically faithful circuits' with a more precise operational term, e.g., 'causally sufficient on target inputs', given the completeness caveats in Appendix E. This would also align the claims with the limitations stated in the Discussion.
  2. [Appendix E, Fig. A3] The redundancy test has two stated conditions (near-zero degradation in the full model, measurable degradation in the circuit), but the figure would be easier to interpret with the decision boundary and the number of red dots in each panel stated explicitly. Also, the toy-model positive control is valuable and should be referenced in the main text.
  3. [§3.2, Fig. 1D] The null distribution is described as 'comparing the first decomposition to an untrained, randomly-initialized version of the second one'. This is a reasonable null, but it would be helpful to also report a permutation null (e.g., matching with randomly permuted component labels) to ensure the match statistics are not inflated by the Hungarian/max-cosine procedure.
  4. [Appendix F] The hyperparameter descriptions are detailed, which is good. Minor notation issue: in Eq. (1), the masks mi are bounded below by μi, but the text later says 'the subcomponent masks are sampled uniformly between mi and μi'—this seems to be a typo for 'between μi and 1'. Please fix.

Circularity Check

1 steps flagged

CSS-only reconstruction is by construction; U-swap is genuine out-of-sample evidence.

specific steps
  1. fitted input called prediction [Section 2 (Method) / Appendix A Eq. 2; Section 3.3 (Extracting a CSS-only submodel)]
    "To achieve this, the ∆ components are adversarially ablated on target batches (each between 0 and 100%) to maximize reconstruction loss, forcing the target outputs to be reconstructed from the subcomponents alone. ... To verify this, we run a “CSS-only” model with only these components on a panel of languages ... As seen in Fig. 2A, the CSS-only model behaves similarly to the original on CSS data (KL-div. ≈ 0.6)."

    The CSS-only model is exactly the target-branch training condition: Eq. (2) adversarially ablates ∆ on target batches so that the subcomponents alone must reproduce the original model's outputs on CSS. Reporting that the CSS-only model matches the original on CSS (KL ≈ 0.6) is therefore checking the objective itself, not independently verifying that the subcomponents are the model's actual CSS mechanisms. The test shows sufficiency of the recovered subcomponents, not faithfulness; it cannot rule out an alternate circuit that also reconstructs CSS. The paper's own Appendix E reinforces this by reporting redundant CSS mechanisms hidden in ∆.

full rationale

The paper is not globally circular: it contains genuine, non-optimized tests. The toy-model ground truth experiments, the seed-consistency and nested-set checks, and especially the 12-block U-swap (Section 3.4) test a counterfactual that the training objective never explicitly optimizes, and the observed np/pd switching is real out-of-sample evidence. However, the CSS-only validation in Section 3.3 is partly by construction: the decomposition was trained so that CSS outputs are reconstructed with ∆ ablated, so the CSS-only model matching the original on CSS is the loss objective's target rather than a test of mechanism identity. The paper itself flags the deeper problem in Section 4 and Appendix E: parameter-decomposition minimality can drop redundant mechanisms, and tPD moves inactive mechanisms into ∆, where they are unobserved. The sentence 'there is preliminary evidence that the CSS-only decomposition misses some redundant mechanisms hidden in the ∆ components' directly undercuts the abstract's 'mechanistically faithful circuits' claim. That means the CSS-only match supports sufficiency, not fidelity, and the abstract overstates what the CSS experiment demonstrates. No self-citation chain or uniqueness theorem is load-bearing; prior PD work is used as framework and benchmark rather than as a circular premise. Overall, one key validation reduces to a fit, while the central editing/swap evidence remains independent, giving partial circularity rather than full reduction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The method depends on several domain assumptions inherited from the PD framework plus the delta residual construction. The most load-bearing is the completeness of Δ: if target-relevant mechanisms are absorbed into Δ, the recovered circuit is incomplete. The paper's own Appendix E indicates this can happen. Free parameters are mostly per-run hyperparameters tuned for each experiment, which weakens the claim of a robust one-configuration method.

free parameters (6)
  • Subcomponent slot counts C per layer = e.g., 96/64 for 4L np/pd; 400/512/126/96/256 for CSS
    Chosen by hand per experiment; determines capacity and FLOPs. No principled selection rule.
  • Importance-minimality coefficient schedule = e.g., 0→4e-3→1e-3 (4L/12L); 0→3e-3→3e-4 (CSS)
    Ad hoc schedule tuned per experiment; directly controls which components survive and hence the final circuit.
  • PGD adversarial loss coefficient and activation time = 0.5 after 80% (4L np/pd); 1 after 20% (12L, CSS)
    Varies across runs; no sensitivity analysis reported.
  • p-norm annealing schedule = 2→1 (np/pd), 2→0.4 (CSS)
    Tuned for stability; no derivation.
  • Unmasked reconstruction loss coefficient (CSS only) = 0.2
    Introduced to prevent dead components from interfering; no ablation justifying this choice.
  • Component matching metric = maximum cosine similarity between flattened UiVj^T
    Choice of metric affects reported agreement; no alternative matching procedures compared.
axioms (4)
  • domain assumption Model computation is faithfully decomposable into rank-1 subcomponents satisfying the four PD requirements (sum to weights, ablatable, simple, minimal activation).
    Foundation of the PD framework; not proven for these models, only empirically validated on toy and small transformer cases.
  • domain assumption Non-target data sampled from the model's original training distribution is representative of all non-target inputs.
    Section 2; if violated, subcomponents remain unconstrained in unseen directions and ablations could have off-target effects on inputs not in the sample.
  • domain assumption Adversarial ablations enforce mechanistic faithfulness, preventing unfaithful alternate implementations.
    Section 2 and Appendix C; relies on the optimization finding the causal mechanisms rather than a different circuit that merely matches outputs.
  • ad hoc to paper The residual Δ = W − Σ UiVi^T captures all non-target mechanisms and no target-relevant mechanisms are lost.
    Eq. (1); Appendix E provides preliminary evidence this is violated (redundant mechanisms absorbed into Δ in the CSS case).
invented entities (1)
  • Catch-all residual component Δ independent evidence
    purpose: Handles all non-target mechanisms without decomposing them; preserves off-target behavior and constrains subcomponents to interfere minimally.
    The paper's naive target-only baseline (no Δ) shows heavy off-target side effects when ablating subcomponents (Fig. 4A), giving Δ empirical justification as a necessary component of the method.

pith-pipeline@v1.3.0-alltime-deepseek · 12860 in / 12970 out tokens · 120317 ms · 2026-08-02T10:35:59.841930+00:00 · methodology

0 comments
read the original abstract

Parameter decomposition (PD) decomposes neural networks into interpretable computational components that faithfully reflect the original network's operations. However, scaling PD to large models requires vast compute, making it a costly and risky endeavor. Here we propose targeted PD (tPD), which identifies only the components that process specific inputs of interest -- from isolated prompts to large subtasks -- by introducing a high-rank catch-all component that handles all non-target data. We validate tPD on toy models and on transformer language models trained on The Pile, where it recovers reproducible, mechanistically faithful circuits. We extract a CSS-only submodel of a 4-block transformer using 7% of the FLOPs of its published decomposition, and in a 12-block transformer we surgically ablate and rewire memorized sequences, with negligible side effects on other inputs.

Figures

Figures reproduced from arXiv: 2607.13047 by Antoine Vigouroux, Lee Sharkey.

Figure 1
Figure 1. Figure 1: A: Causal importance patterns for the TMCC toy model’s first matrix: full-data and targeted on 3 features. B: Mean-square error (MSE) of the reconstruction, as a proxy for decomposition success, for targeted decomposition of the TMCC toy model when a variable number of subcomponent slots are provided. C: MSE over training for the full-data and targeted decompositions of the TMCC. D: Cosine similarities of … view at source ↗
Figure 2
Figure 2. Figure 2: A: KL-divergence of a model made only of the sub￾components identified in the CSS-only decomposition, compared to the original model, on various languages. Lower values mean the outputs are closer to the original model’s. B: Comparison of the original sequence, next tokens predicted by the original model, and next tokens predicted by the CSS-only model, on sam￾ples of text from CSS code, Python code, and E… view at source ↗
Figure 3
Figure 3. Figure 3: Three examples of CSS-specific subcomponents from the CSS-only decomposition. Left: KL-divergence compared to the original model’s output when the subcomponent is individually subtracted from the model, for CSS code versus other languages. Right: Activation patterns of these components (displayed as the per-token KL-divergence when ablated, clamped to [0; 1]). Using ≈7% of the FLOPs of the published full-d… view at source ↗
Figure 4
Figure 4. Figure 4: A: Effect of ablating individual subcomponents obtained either with the tPD method introduced here (“targeted”), or with standard PD on the target data without non-target batches (“naive”). The x- and y-axes show the maximum KL(original||ablated) over the two target sequences. Point color represents the 99th-percentile KL-divergence on >100k tokens of general Pile data. Higher values mean the ablated model… 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

174 extracted references · 86 canonical work pages · 3 internal anchors

  1. [1]

    Activation space interpretability may be doomed —

    bilalchughtai and Bushnaq, Lucius , month = jan, year =. Activation space interpretability may be doomed —

  2. [2]

    Transformer

    Geva, Mor and Schuster, Roei and Berant, Jonathan and Levy, Omer , month = sep, year =. Transformer. doi:10.48550/arXiv.2012.14913 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2012.14913 2012

  3. [3]

    Pan, Haowen and Wang, Xiaozhi and Cao, Yixin and Shi, Zenglin and Yang, Xun and Li, Juanzi and Wang, Meng , month = mar, year =. Precise. doi:10.48550/arXiv.2503.01090 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.01090

  4. [4]

    EvoEdit: Evolving Null-space Alignment for Robust and Efficient Knowledge Editing

    Lyu, Sicheng and Gu, Yu and Wang, Xinyu and Huang, Jerry and Luan, Sitao and Cui, Yufei and Chang, Xiao-Wen and Lu, Peng , month = apr, year =. doi:10.48550/arXiv.2510.13851 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.13851

  5. [5]

    and Andreas, Jacob , month = aug, year =

    Hernandez, Evan and Li, Belinda Z. and Andreas, Jacob , month = aug, year =. Inspecting and. doi:10.48550/arXiv.2304.00740 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.00740

  6. [6]

    SAKE: Steering Activations for Knowledge Editing

    Scialanga, Marco and Laugel, Thibault and Grari, Vincent and Detyniecki, Marcin , month = jul, year =. doi:10.48550/arXiv.2503.01751 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.01751

  7. [7]

    Reinforced

    Li, Zherui and Jiang, Houcheng and Chen, Hao and Bi, Baolong and Zhou, Zhenhong and Sun, Fei and Fang, Junfeng and Wang, Xiang , month = sep, year =. Reinforced. doi:10.48550/arXiv.2502.05759 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.05759

  8. [8]

    doi:10.48550/arXiv.2410.02355 , abstract =

    Fang, Junfeng and Jiang, Houcheng and Wang, Kun and Ma, Yunshan and Jie, Shi and Wang, Xiang and He, Xiangnan and Chua, Tat-seng , month = apr, year =. doi:10.48550/arXiv.2410.02355 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.02355

  9. [9]

    McGrath, Thomas and Rahtz, Matthew and Kramar, Janos and Mikulik, Vladimir and Legg, Shane , month = jul, year =. The. doi:10.48550/arXiv.2307.15771 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.15771

  10. [10]

    Interpreting

    Bushnaq, AUTHORS Lucius and Braun, * Dan and. Interpreting

  11. [11]

    Transcoders

    Dunefsky, Jacob and Chlenski, Philippe and Nanda, Neel , month = nov, year =. Transcoders. doi:10.48550/arXiv.2406.11944 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.11944

  12. [12]

    Distill , author =

    High/. Distill , author =. 2021 , pages =. doi:10.23915/distill.00024.005 , number =

    doi:10.23915/distill.00024.005 2021

  13. [13]

    Distill , author =

    Multimodal. Distill , author =. 2021 , pages =. doi:10.23915/distill.00030 , number =

    doi:10.23915/distill.00030 2021

  14. [14]

    Bau, David and Zhou, Bolei and Khosla, Aditya and Oliva, Aude and Torralba, Antonio , month = jul, year =. Network. 2017. doi:10.1109/CVPR.2017.354 , urldate =

    doi:10.1109/cvpr.2017.354 2017

  15. [15]

    Distill , author =

    Zoom. Distill , author =. 2020 , pages =. doi:10.23915/distill.00024.001 , number =

    doi:10.23915/distill.00024.001 2020

  16. [16]

    He, Zhengfu and Wang, Junxuan and Lin, Rui and Ge, Xuyang and Shu, Wentao and Tang, Qiong and Zhang, Junping and Qiu, Xipeng , month = apr, year =. Towards. doi:10.48550/arXiv.2504.20938 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.20938

  17. [17]

    OpenMOSS Aiology Posts , author =

    Bridging the. OpenMOSS Aiology Posts , author =

  18. [18]

    Decoding

    Muhamed, Aashiq and Diab, Mona and Smith, Virginia , month = nov, year =. Decoding. doi:10.48550/arXiv.2411.00743 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.00743

  19. [19]

    Resurrecting the Salmon: Rethinking Mechanistic Interpretability with Domain-Specific Sparse Autoencoders

    O'Neill, Charles and Jayasekara, Mudith and Kirkby, Max , month = aug, year =. Resurrecting the. doi:10.48550/arXiv.2508.09363 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.09363

  20. [20]

    How to use and interpret activation patching , url =

    Heimersheim, Stefan and Nanda, Neel , month = apr, year =. How to use and interpret activation patching , url =. doi:10.48550/arXiv.2404.15255 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.15255

  21. [21]

    Locating and

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , month = jan, year =. Locating and. doi:10.48550/arXiv.2202.05262 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2202.05262

  22. [22]

    and Zhou, Georgia and Cherapanamjeri, Yeshwanth and Huang, Yaxuan and Odisho, Anobel Y

    Hsu, Aliyah R. and Zhou, Georgia and Cherapanamjeri, Yeshwanth and Huang, Yaxuan and Odisho, Anobel Y. and Carroll, Peter R. and Yu, Bin , month = mar, year =. Efficient. doi:10.48550/arXiv.2407.00886 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.00886

  23. [23]

    Sharkey, Lee and Chughtai, Bilal and Batson, Joshua and Lindsey, Jack and Wu, Jeff and Bushnaq, Lucius and Goldowsky-Dill, Nicholas and Heimersheim, Stefan and Ortega, Alejandro and Bloom, Joseph and Biderman, Stella and Garriga-Alonso, Adria and Conmy, Arthur and Nanda, Neel and Rumbelow, Jessica and Wattenberg, Martin and Schoots, Nandi and Miller, Jose...

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.16496

  24. [24]

    Meng, Kevin and Sharma, Arnab Sen and Andonian, Alex and Belinkov, Yonatan and Bau, David , month = aug, year =. Mass-. doi:10.48550/arXiv.2210.07229 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.07229

  25. [25]

    and Lynch, Aengus and Heimersheim, Stefan and Garriga-Alonso, Adrià , month = oct, year =

    Conmy, Arthur and Mavor-Parker, Augustine N. and Lynch, Aengus and Heimersheim, Stefan and Garriga-Alonso, Adrià , month = oct, year =. Towards. doi:10.48550/arXiv.2304.14997 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.14997

  26. [26]

    Cunningham, Hoagy and Ewart, Aidan and Riggs, Logan and Huben, Robert and Sharkey, Lee , month = oct, year =. Sparse. doi:10.48550/arXiv.2309.08600 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.08600

  27. [27]

    Kocetkov, Denis and Li, Raymond and Allal, Loubna Ben and Li, Jia and Mou, Chenghao and Ferrandis, Carlos Muñoz and Jernite, Yacine and Mitchell, Margaret and Hughes, Sean and Wolf, Thomas and Bahdanau, Dzmitry and Werra, Leandro von and Vries, Harm de , month = nov, year =. The. doi:10.48550/arXiv.2211.15533 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.15533

  28. [28]

    Merity, Stephen and Xiong, Caiming and Bradbury, James and Socher, Richard , month = sep, year =. Pointer. doi:10.48550/arXiv.1609.07843 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1609.07843

  29. [29]

    , month = jun, year =

    Mitchell, Eric and Lin, Charles and Bosselut, Antoine and Finn, Chelsea and Manning, Christopher D. , month = jun, year =. Fast. doi:10.48550/arXiv.2110.11309 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2110.11309

  30. [30]

    Eldan, Ronen and Russinovich, Mark , month = oct, year =. Who's. doi:10.48550/arXiv.2310.02238 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.02238

  31. [31]

    Knowledge

    Dai, Damai and Dong, Li and Hao, Yaru and Sui, Zhifang and Chang, Baobao and Wei, Furu , editor =. Knowledge. Proceedings of the 60th. 2022 , pages =. doi:10.18653/v1/2022.acl-long.581 , abstract =

    doi:10.18653/v1/2022.acl-long.581 2022

  32. [32]

    Interpretability in

    Braun, Dan and Bushnaq, Lucius and Heimersheim, Stefan and Mendel, Jake and Sharkey, Lee , month = feb, year =. Interpretability in. doi:10.48550/arXiv.2501.14926 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.14926

  33. [33]

    and Riggs, Logan , month = dec, year =

    Christensen, Casper L. and Riggs, Logan , month = dec, year =. Decomposition of. doi:10.48550/arXiv.2511.08854 , abstract =

    work page doi:10.48550/arxiv.2511.08854

  34. [34]

    Interpretability in the

    Wang, Kevin and Variengien, Alexandre and Conmy, Arthur and Shlegeris, Buck and Steinhardt, Jacob , month = nov, year =. Interpretability in the. doi:10.48550/arXiv.2211.00593 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.00593

  35. [35]

    Stochastic

    Bushnaq, Lucius and Braun, Dan and Sharkey, Lee , month = sep, year =. Stochastic. doi:10.48550/arXiv.2506.20790 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.20790

  36. [36]

    Making sense of parameter-space decomposition —

    Malmesbury , month = sep, year =. Making sense of parameter-space decomposition —

  37. [37]

    System drift and speciation

  38. [38]

    Journal of Bacteriology , author =

    Escherichia coli. Journal of Bacteriology , author =. 2007 , pages =. doi:10.1128/JB.01368-07 , abstract =

    work page doi:10.1128/jb.01368-07 2007

  39. [39]

    1983 , pages =

    Evolution , author =. 1983 , pages =. doi:10.1111/j.1558-5646.1983.tb05521.x , number =

    arXiv 1983

  40. [40]

    Molecular and General Genetics MGG , author =

    Selection against hypermutability in. Molecular and General Genetics MGG , author =. 1984 , keywords =. doi:10.1007/BF00328720 , abstract =

    work page doi:10.1007/bf00328720 1984

  41. [41]

    and Sniegowski, P

    Raynes, Y. and Sniegowski, P. D. , month = nov, year =. Experimental evolution and the dynamics of genomic mutation rate modifiers , volume =. Heredity , publisher =. doi:10.1038/hdy.2014.49 , abstract =

    work page doi:10.1038/hdy.2014.49 2014

  42. [42]

    1996 , keywords =

    The EMBO journal , author =. 1996 , keywords =

    1996

  43. [44]

    Proceedings of the National Academy of Sciences of the United States of America , author =

    Improved bacterial recombineering by parallelized protein discovery , volume =. Proceedings of the National Academy of Sciences of the United States of America , author =. 2020 , pages =. doi:10.1073/pnas.2001588117 , abstract =

    work page doi:10.1073/pnas.2001588117 2020

  44. [45]

    BMC Biotechnology , author =

    Long term adaptation of a microbial population to a permanent metabolic constraint: overcoming thymineless death by experimental evolution of. BMC Biotechnology , author =. 2001 , keywords =. doi:10.1186/1472-6750-1-10 , abstract =

    work page doi:10.1186/1472-6750-1-10 2001

  45. [46]

    Microbial Biotechnology , author =

    An automated. Microbial Biotechnology , author =. 2021 , note =. doi:10.1111/1751-7915.13678 , abstract =

    arXiv 2021

  46. [47]

    Ecology and Evolution , author =

    Temperature affects the repeatability of evolution in the microbial eukaryote. Ecology and Evolution , author =. 2021 , note =. doi:10.1002/ece3.8036 , abstract =

    work page doi:10.1002/ece3.8036 2021

  47. [48]

    Genome Biology and Evolution , author =

    Evolutionary. Genome Biology and Evolution , author =. 2021 , pages =. doi:10.1093/gbe/evab257 , abstract =

    work page doi:10.1093/gbe/evab257 2021

  48. [49]

    Journal of Molecular Evolution , author =

    Revisiting the. Journal of Molecular Evolution , author =. 2023 , keywords =. doi:10.1007/s00239-023-10095-3 , abstract =

    work page doi:10.1007/s00239-023-10095-3 2023

  49. [50]

    Molecular Systems Biology , author =

    Construction of. Molecular Systems Biology , author =. 2006 , keywords =. doi:10.1038/msb4100050 , abstract =

    work page doi:10.1038/msb4100050 2006

  50. [51]

    G3 Genes

    Zheng, Qi , month = dec, year =. G3 Genes. doi:10.1534/g3.117.300120 , abstract =

    work page doi:10.1534/g3.117.300120

  51. [52]

    Methods in Molecular Biology (Clifton, N.J.) , author =

    Measuring. Methods in Molecular Biology (Clifton, N.J.) , author =. 2018 , keywords =. doi:10.1007/978-1-4939-7306-4_3 , abstract =

    work page doi:10.1007/978-1-4939-7306-4_3 2018

  52. [53]

    Exploring the fitness benefits of genome reduction in

    Vernyik, Viktor and Karcagi, Ildikó and Tímár, Edit and Nagy, István and Györkei, Ádám and Papp, Balázs and Györfy, Zsuzsanna and Pósfai, György , month = apr, year =. Exploring the fitness benefits of genome reduction in. Scientific Reports , publisher =. doi:10.1038/s41598-020-64074-5 , abstract =

    work page doi:10.1038/s41598-020-64074-5

  53. [54]

    PLoS Genetics , author =

    Repeated,. PLoS Genetics , author =. 2012 , pages =. doi:10.1371/journal.pgen.1002651 , abstract =

    work page doi:10.1371/journal.pgen.1002651 2012

  54. [55]

    Metabolic efficiency and amino acid composition in the proteomes of

    Akashi, Hiroshi and Gojobori, Takashi , month = mar, year =. Metabolic efficiency and amino acid composition in the proteomes of. Proceedings of the National Academy of Sciences , publisher =. doi:10.1073/pnas.062526999 , abstract =

    work page doi:10.1073/pnas.062526999

  55. [56]

    The ISME journal , author =

    Genome reduction occurred in early. The ISME journal , author =. 2024 , keywords =. doi:10.1093/ismejo/wrad035 , abstract =

    work page doi:10.1093/ismejo/wrad035 2024

  56. [57]

    and Arkin, Adam P

    Price, Morgan N. and Arkin, Adam P. , month = dec, year =. Weakly. mBio , publisher =. doi:10.1128/mbio.01302-15 , abstract =

    work page doi:10.1128/mbio.01302-15

  57. [58]

    Science (New York, N.Y.) , author =

    The origins of genome complexity , volume =. Science (New York, N.Y.) , author =. 2003 , keywords =. doi:10.1126/science.1089370 , abstract =

    work page doi:10.1126/science.1089370 2003

  58. [59]

    The unexhausted potential of

    Blount, Zachary D , month = mar, year =. The unexhausted potential of. eLife , publisher =. doi:10.7554/eLife.05826 , abstract =

    work page doi:10.7554/elife.05826

  59. [60]

    Proceedings of the Royal Society B: Biological Sciences , author =

    The distribution of bacterial doubling times in the wild , volume =. Proceedings of the Royal Society B: Biological Sciences , author =. 2018 , pages =. doi:10.1098/rspb.2018.0789 , abstract =

    arXiv 2018

  60. [61]

    , month = dec, year =

    Lynch, Michael and Marinov, Georgi K. , month = dec, year =. The bioenergetic costs of a gene , volume =. Proceedings of the National Academy of Sciences , publisher =. doi:10.1073/pnas.1514974112 , abstract =

    work page doi:10.1073/pnas.1514974112

  61. [62]

    Proceedings of the National Academy of Sciences of the United States of America , author =

    Theory of prokaryotic genome evolution , volume =. Proceedings of the National Academy of Sciences of the United States of America , author =. 2016 , pages =. doi:10.1073/pnas.1614083113 , abstract =

    work page doi:10.1073/pnas.1614083113 2016

  62. [63]

    Trends in genetics: TIG , author =

    Deletional bias and the evolution of bacterial genomes , volume =. Trends in genetics: TIG , author =. 2001 , keywords =. doi:10.1016/s0168-9525(01)02447-7 , abstract =

    work page doi:10.1016/s0168-9525(01)02447-7 2001

  63. [64]

    Vieira-Silva, Sara and Touchon, Marie and Rocha, Eduardo P. C. , month = jun, year =. No evidence for elemental-based streamlining of prokaryotic genomes , volume =. Trends in Ecology & Evolution , publisher =. doi:10.1016/j.tree.2010.03.001 , language =

    work page doi:10.1016/j.tree.2010.03.001 2010

  64. [65]

    Genome Biology and Evolution , author =

    Core. Genome Biology and Evolution , author =. 2017 , pages =. doi:10.1093/gbe/evx064 , abstract =

    work page doi:10.1093/gbe/evx064 2017

  65. [66]

    , month = oct, year =

    Bremer, Hans and Dennis, Patrick P. , month = oct, year =. Modulation of. EcoSal Plus , publisher =. doi:10.1128/ecosal.5.2.3 , abstract =

    work page doi:10.1128/ecosal.5.2.3

  66. [67]

    2025 , keywords =

    Nucleic Acids Research , author =. 2025 , keywords =. doi:10.1093/nar/gkae1244 , abstract =

    work page doi:10.1093/nar/gkae1244 2025

  67. [68]

    Journal of Bacteriology , author =

    Chromosomal locations of the genes for. Journal of Bacteriology , author =. 1982 , pages =. doi:10.1128/jb.149.2.458-468.1982 , abstract =

    work page doi:10.1128/jb.149.2.458-468.1982 1982

  68. [70]

    DNA repair , author =

    Replication fidelity in. DNA repair , author =. 2019 , pages =. doi:10.1016/j.dnarep.2019.102643 , abstract =

    arXiv 2019

  69. [71]

    Speed variations of bacterial replisomes , volume =

    Bhat, Deepak and Hauf, Samuel and Plessy, Charles and Yokobayashi, Yohei and Pigolotti, Simone , editor =. Speed variations of bacterial replisomes , volume =. eLife , publisher =. 2022 , keywords =. doi:10.7554/eLife.75884 , abstract =

    work page doi:10.7554/elife.75884 2022

  70. [72]

    The possible fidelity-speed-proofreading cost trade-offs in

    Li, Qiu-Shi and Shu, Yao-Gen and Fu, Wen-Bo and Ou-Yang, Zhong-Can and Li, Ming , month = feb, year =. The possible fidelity-speed-proofreading cost trade-offs in. doi:10.1101/2021.02.18.431768 , abstract =

    work page doi:10.1101/2021.02.18.431768 2021

  71. [73]

    Journal of Molecular Biology , author =

    Chromosome replication and the division cycle of. Journal of Molecular Biology , author =. 1968 , pages =. doi:10.1016/0022-2836(68)90425-7 , abstract =

    work page doi:10.1016/0022-2836(68)90425-7 1968

  72. [74]

    The EMBO Journal , author =

    Timing of initiation of chromosome replication in individual. The EMBO Journal , author =. 1986 , pages =. doi:10.1002/j.1460-2075.1986.tb04415.x , abstract =

    arXiv 1986

  73. [75]

    , month = nov, year =

    Gon, Stéphanie and Napolitano, Rita and Rocha, Walter and Coulon, Stéphane and Fuchs, Robert P. , month = nov, year =. Increase in. Proceedings of the National Academy of Sciences , publisher =. doi:10.1073/pnas.1113664108 , abstract =

    work page doi:10.1073/pnas.1113664108

  74. [76]

    DNA repair , author =

    Novel mutator mutants of. DNA repair , author =. 2012 , keywords =. doi:10.1016/j.dnarep.2012.02.001 , abstract =

    work page doi:10.1016/j.dnarep.2012.02.001 2012

  75. [77]

    The Journal of Biological Chemistry , author =

    The polymerase subunit of. The Journal of Biological Chemistry , author =. 1985 , keywords =

    1985

  76. [78]

    Molecular Microbiology , author =

    Ribonucleotide reductase and the regulation of. Molecular Microbiology , author =. 2007 , note =. doi:10.1111/j.1365-2958.2006.05493.x , abstract =

    arXiv 2007

  77. [79]

    Odsbu, Ingvild and Morigen and Skarstad, Kirsten , month = oct, year =. A. PLOS ONE , publisher =. doi:10.1371/journal.pone.0007617 , abstract =

    work page doi:10.1371/journal.pone.0007617

  78. [80]

    and Semsey, Szabolcs and Gerdes, Kenn , month = dec, year =

    Harms, Alexander and Fino, Cinzia and Sørensen, Michael A. and Semsey, Szabolcs and Gerdes, Kenn , month = dec, year =. Prophages and. mBio , publisher =. doi:10.1128/mbio.01964-17 , abstract =

    work page doi:10.1128/mbio.01964-17

  79. [81]

    , year =

    Vasi, Farida and Travisano, Michael and Lenski, Richard E. , year =. Long-. The American Naturalist , publisher =

  80. [82]

    Journal of Genetics , author =

    Ecological strategies and fitness tradeoffs. Journal of Genetics , author =. 1999 , keywords =. doi:10.1007/BF02994702 , abstract =

    work page doi:10.1007/bf02994702 1999

Showing first 80 references.