Pith. sign in

REVIEW 4 major objections 5 minor 70 references

On the Mechanistic Interpretability of Neural Networks for Causality in Bio-statistics

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Mechanistic interpretability tools—probing, ablation, and causal tracing—can open up the neural networks used inside TMLE causal estimators, showing whether a network encodes known confounders and which input pathways drive its estimates.

desk verdict Reproducible probe/ablation proof of concept undermined by an unsupported treatment-pathway claim. read the letter →

arxiv 2505.00555 v1 pith:OQDKP4VR submitted 2025-05-01 stat.AP cs.AI

Jean-Baptiste A. Conan This is my paper · ORCID
classification stat.APcs.AI MSC 62D2068T07
keywords mechanisticinterpretabilitycausalinferenceTMLEnuisancefunctionestimationlinearprobingtracingneuralnetworkconfounding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

In bio-statistics, causal effect estimates increasingly rely on neural networks to estimate nuisance functions inside Targeted Minimum Loss-based Estimation (TMLE), but these networks are hard to trust because their internals are opaque. This paper argues that mechanistic interpretability tools developed for large language models transfer to these smaller networks: linear probes can detect whether a known confounder is encoded in the shared layers, ablating the neurons that encode it moves the final average treatment effect estimate, and causal tracing can draw the computational pathways through which inputs such as confounders and treatments propagate. The experiments are carried out on synthetic data with a known true treatment effect, so the internal findings can be checked against ground truth. If the demonstration holds, MI gives biostatisticians a way to validate and debug network-based nuisance estimators before relying on their causal conclusions.

What carries the argument

Three interacting tools carry the argument. Linear probing: train a simple regression on activations at each shared layer to predict the known confounder $W_1$, then rank neurons by the magnitude of the probe weights; high probe $R^2$ is evidence the confounder is represented. Neuron ablation: zero the output of the probe-identified neurons during a forward pass and measure how much the $Q$- and $g$-head losses and the TMLE ATE estimate move; larger movement for top-importance neurons is evidence the representation is causally used. Causal tracing: patch the activation of one input neuron, then successively patch downstream neurons whose activation change exceeds a cutoff, building a directed graph of 'impacted' nodes; sparsity, success, and Jaccard-overlap scores turn these graphs into comparable pathway summaries.

What would settle it

Re-run the causal-tracing procedure of Section 5.2 on the same trained networks while sweeping the activation-change cutoff across a plausible range (e.g., from a small fixed value to a high percentile of the null distribution of activation differences induced by patching). If the reported treatment-versus-confounder pathway separation, the success scores, and the Jaccard-overlap structure change qualitatively across cutoffs, the discovered pathways are threshold artifacts. A second check: train the same architecture on a version of DS1 in which the confounder $W_1$ is permuted to break its link to treatment and outcome; if probes still report high $R^2$ for $W_1$ or ablating probe-important neurons still moves the TMLE ATE, the representation-and-ablation claims fail.

Watch

Extended reading notes

Core claim

The central claim is that MI techniques—linear probing, neuron ablation, and causal tracing—can be applied to neural networks that estimate nuisance functions in TMLE, and that they reveal something true about how those networks compute. Concretely, the paper shows that a linear probe trained on shared-layer activations recovers the critical confounder $W_1$ with high $R^2$ at every layer; that ablating the neurons most important to that probe shifts the resulting TMLE estimate of the average treatment effect more than ablating random or least-important neurons; and that causal traces started from individual input neurons produce pathway graphs with measurable overlap structure, which the author interprets as evidence that the network processes confounder and treatment information through partly distinct internal routes. The ground truth for these experiments comes from synthetic datasets with known true ATEs (2.0 in the strong-confounder design, 0 in the no-effect design).

Load-bearing premise

The load-bearing premise is that the causal-tracing pathway graphs are not artifacts of an unstated activation-change cutoff: Section 5.2.1 marks downstream neurons as 'failed' whenever their activation change does not exceed a threshold, but the threshold is never specified, so the existence of distinct confounder and treatment pathways depends on an untested tuning choice.

Editorial extensions

If this is right

  • Linear probing can serve as a validation check: a TMLE nuisance network that linearly encodes a known confounder passes a minimal test that it has learned the right input structure.
  • Ablation of probe-important neurons changes the final ATE estimate, so the internal representation is not a correlation artifact—it is load-bearing for the causal estimate.
  • Causal tracing provides a visual and quantitative map of how individual input covariates propagate; treatment and confounder inputs can be compared by pathway overlap rather than by inspecting weights.
  • Because the same tools work on any network with shared layers, the approach extends to other nuisance estimators and to longitudinal variants of TMLE.
  • If MI validation becomes standard for network-based nuisance estimation, black-box concerns in high-stakes biostatistics shift from 'can we trust it' to 'which confounders are demonstrably used and how'.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable next step is to turn the pathway-overlap matrix into a formal diagnostic: under a null model with randomized weights or shuffled inputs, the null distribution of Jaccard overlap between treatment and confounder traces could be estimated, so a reported separation can be tested rather than read off a graph.
  • If linear decodability and ablation-sensitivity generalize to real electronic health record data, the same protocol could be used as a pre-deployment audit for any neural nuisance model, flagging networks that encode confounders they should not or fail to encode confounders they must adjust for.
  • The probe-importance ranking also suggests a pruning or regularization strategy: neurons that are important for a confounder could be penalized or removed when the goal is to estimate effects invariant to that confounder, though this is a design choice the paper does not explore.
  • The author's framing implies a broader claim: MI is not only for frontier models but can serve as a general science-of-models toolkit for statistics, where ground truth about causal structure is often known from the study design.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This manuscript, a master's thesis, proposes applying mechanistic interpretability (MI) tools to neural networks used for causal inference in biostatistics, specifically multi-task networks that estimate nuisance functions within TMLE. In Experiment 1, linear probes are trained on shared-layer activations to predict a known strong confounder W1, and ablation of the probe-selected neurons is shown to shift the resulting TMLE ATE estimate relative to random and bottom-neuron ablations. Experiments 2 and 3 apply causal tracing from input neurons to generate pathway graphs and Jaccard-overlap matrices, which the paper interprets as revealing distinct computational pathways for confounders versus treatments. The paper concludes that MI can validate nuisance function estimators and reveal how different input covariates are processed, and it claims a third contribution of providing methods to compare mechanisms across statistical, ML, and NN models.

Significance. The work is a useful proof-of-concept that linear probing and ablation can provide external validation of nuisance networks: the experiments use synthetic data with a known true ATE, compare against random and bottom-neuron ablation baselines, and the code is publicly referenced. If the causal-tracing claims were fully supported, the paper would make a modest but real contribution to the emerging practice of auditing black-box causal estimators. The primary experimental evidence for claim (1) is credible; the evidence for claim (2) is not, because treatment information is never patched, and claim (3) is not experimentally addressed.

major comments (4)
  1. [§5.2.1, §5.3.1, Chapter 6] The causal-tracing experiments patch only input neurons corresponding to covariates W, never the treatment input A; because in fQ,g the treatment A enters only the Q-head (fQ(h_shared,A)) while the shared layers process W, the traced pathways describe how covariates propagate and cannot support the conclusion that the network has distinct 'treatment vs. confounder' pathways. The claim in Chapter 6 that tracing revealed 'how information from different input covariates, such as confounders versus treatment variables, is processed' overreaches the collected data; a treatment-input patch or an explicit head-level intervention would be needed.
  2. [§5.2.1] The method marks nodes as 'failed' when activation changes do not exceed a threshold, but the threshold is never specified and no sensitivity analysis is reported; because the success score, the edge set in Figures 5.7-5.9, and all overlap computations depend on this threshold, the pathway graphs are not reproducible or interpretable without reporting it.
  3. [Abstract and §5.1-5.3] The abstract's third claimed contribution—providing methodologies for comparing learned mechanisms across statistical, ML, and NN models—is not realized anywhere in the experiments; Sections 5.1-5.3 study only the NN fQ,g, with no cross-model comparison, and Chapter 6 does not present such a method.
  4. [§5.1.1 and Appendix 'Experiment 1'] The architecture description is internally inconsistent: §5.1.1 and Table 5.1 refer to shared layers h1 through h8 plus h_shared, while the architecture section states 5 shared layers with 100 units, and the training section reports that 'HIDDEN LAYER = 10 and HIDDEN SIZE = 30' gave the best results; the actual network used in Experiments 1 and 3 is therefore ambiguous, preventing reproducibility.
minor comments (5)
  1. [Chapter 5 opening] Chapter 5 opens with 'this chapter presents experiments1 applying MI techniques' and refers to 'the previous chapter ??'; fix the broken cross-reference.
  2. [§5.2.2 and §5.3.2] Figure 5.6 is described in §5.2.2 as the no-effect model and in §5.3.2 as the strong-confounder model; the caption and text should state which dataset the figure corresponds to.
  3. [Appendix 'All graphs from Experience 3 (5.2)'] The appendix heading 'All graphs from Experience 3 (5.2)' is mislabeled; the figures it contains appear to belong to the Section 5.2 causal-tracing experiments.
  4. [§5.2.1] Section 5.2.1 contains the typo 'follwi metrics,' and the Success Score definition depends on the unspecified threshold noted in the major comments.
  5. [§5.1.1 and Appendix 'Training'] The combined loss in Section 5.1.1 uses weights wq and wg, while the appendix reports a single alpha = 0.5; specify the relationship between these quantities.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the experiments are benchmarked against synthetic ground truth and controlled ablations, and the main overreach is a scope gap rather than a derivation from inputs.

full rationale

The paper's central derivation chain is externally anchored. In Experiment 1, the TMLE nuisance network is trained on synthetic DS1 with a known true ATE of 2.0, and the ablated-network ATE estimates are compared against this ground truth; this is an external benchmark, not a fitted parameter renamed as a prediction. The probe selects neurons by their linear predictiveness of the confounder W1, but the subsequent ablation is an independent intervention, and the paper includes Bottom-10% and Random-10% ablation controls that reduce the risk that the result is forced by the selection procedure. In Experiments 2 and 3, causal tracing produces pathway visualizations defined by activation-patch impact; these are descriptive outputs of an interventional method, and no quantitative claim is made that reduces to the method's own threshold by construction. The unspecified activation-change threshold is a reproducibility and robustness concern, not circularity. The conclusion that the network has distinct 'treatment vs. confounder pathways' is not actually supported because the treatment input A enters only at the Q-head and is never patched as an input neuron, but that is an overclaim about experimental scope, not a circular derivation. Citations to van der Laan and TMLE are background methodology, not a load-bearing self-citation chain or an imported uniqueness theorem. Therefore no circular step meets the evidentiary bar of the review rules.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on standard ML training choices plus several domain-specific interpretability assumptions: that linear probe accuracy is a valid gauge of representation, that ablation of probe-important neurons identifies causal reliance, and that thresholded causal-tracing graphs mirror real computation. No new physical or conceptual entities are postulated. The critical unspecified quantity is the causal-tracing activation-change threshold.

free parameters (3)
  • Combined loss weight alpha = 0.5
    Balances MSE (Q-head) and BCE (g-head) in ToyTMLELoss (Appendix, Experiment 1 Training). It shapes the learned representations and the ablation results, but is not fitted to force the target claim.
  • Model architecture (number and size of hidden layers) = contradictory: 5 layers of 100 units vs 10 layers of 30 units
    The appendix architecture definition says 5 shared layers of 100 units, while the training section reports that 10 layers of 30 units 'yielded the best results'. This inconsistency changes which activations are probed in Table 5.1 and affects all experimental results.
  • Causal-tracing activation-change threshold
    Pathway propagation and node 'failure' depend on an activation-change threshold that is never defined. This parameter determines the sparsity score, success score, and all pathway figures.
assumptions (4)
  • domain assumption Linear probe accuracy is a valid measure of how strongly a concept is represented in a layer.
    Probe results in Section 5.1.2 are interpreted as 'W1 is strongly and linearly represented'. This assumes linear decodability reflects a meaningful internal representation rather than an artifact of the probe or data distribution.
  • domain assumption Ablating probe-important neurons and observing output changes identifies causal reliance on the probed concept.
    The ablation analysis attributes ATE shifts to loss of W1 information, but any functionally important neuron ablation could perturb the estimate. Random and bottom-neuron controls only partially address this.
  • domain assumption Thresholded causal-tracing graphs faithfully represent the network's computational pathways.
    The tracing method in Sections 5.2 and 5.3 uses an unspecified threshold to decide whether a downstream activation change is significant; the resulting graphs are treated as evidence of real computational pathways.
  • domain assumption Synthetic data with known ATE is an adequate benchmark for validating NN nuisance estimators in TMLE.
    Experiments use only simple simulated DAGs (DS1 and DS2) with known truth. The paper does not test on realistic high-dimensional or real-world data, so the benchmark is idealized.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Mechanistic Interpretability of Neural Networks for Causality in Bio-statistics." pith.science (2026). https://pith.science/paper/OQDKP4VR

@misc{pith2026250500555,
  author       = {Pith},
  title        = {Pith review of: On the Mechanistic Interpretability of Neural Networks for Causality in Bio-statistics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQDKP4VR}},
  note         = {Machine review of arXiv:2505.00555}
}
read the original abstract

Interpretable insights from predictive models remain critical in bio-statistics, particularly when assessing causality, where classical statistical and machine learning methods often provide inherent clarity. While Neural Networks (NNs) offer powerful capabilities for modeling complex biological data, their traditional "black-box" nature presents challenges for validation and trust in high-stakes health applications. Recent advances in Mechanistic Interpretability (MI) aim to decipher the internal computations learned by these networks. This work investigates the application of MI techniques to NNs within the context of causal inference for bio-statistics. We demonstrate that MI tools can be leveraged to: (1) probe and validate the internal representations learned by NNs, such as those estimating nuisance functions in frameworks like Targeted Minimum Loss-based Estimation (TMLE); (2) discover and visualize the distinct computational pathways employed by the network to process different types of inputs, potentially revealing how confounders and treatments are handled; and (3) provide methodologies for comparing the learned mechanisms and extracted insights across statistical, machine learning, and NN models, fostering a deeper understanding of their respective strengths and weaknesses for causal bio-statistical analysis.

Figures

Figures reproduced from arXiv: 2505.00555 by the authors.

Figure 1.1
Figure 1.1. Visualization of AI model scaling trends based on the Epoch AI Notable Models dataset [Epoch AI, 2024]. The figure illustrates key insights including: training compute doubling every five months since 2010 (growing at 4.7x per year), training costs for large models doubling every nine months, dataset sizes for language models doubling every eight months, and power requirements for frontier models doubling annually. … view at source ↗
Figure 1.2
Figure 1.2. Estimated Average Per-Patient Clinical Trial Costs by Selected Condition (Battelle Memorial Institute [2015]) 1.1.3 Statistics to optimize Clinical Trials Sampling size and statistical significance Optimizing clinical trials involves strategically determining the minimal sample size required to achieve statistical significance, thereby reducing costs without compromising result validity. Statistical significance ass… view at source ↗
Figure 2.1
Figure 2.1. Tokenization of the sentence ”For example, the current sentence’s tokenization is shown in [PITH_FULL_IMAGE:figures/full_fig_p034_2_1.png] view at source ↗
Figures from the paper (19 more)
Figure 5.1
Figure 5.1. Figure 5.1: Probe R² and Neuron Counts for Importance Thresholds Across Layers. This plot shows the R² score (blue line, left axis) of linear probes predicting W1 from each shared layer’s activations. The red lines (right axis) show the number of neurons (sorted by probe coeffic…
Figure 5.2
Figure 5.2. Figure 5.2: Cumulative Importance vs. % Neurons Across Layers. Each curve represents a shared layer, showing the percentage of total W1 importance (sum of absolute probe coefficients) captured as more neurons (sorted by importance) are included. Steeper curves indicate higher co…
Figure 5.3
Figure 5.3. Figure 5.3: Experiment 1: Absolute ATE (10% Neurons Ablated). This bar chart compares the TMLE ATE estimate (with 95% CI) obtained using the original network outputs (Baseline ’TMLE’) versus outputs after ablating the Top 10%, Bottom 10%, or Random 10% most important neurons (ac…
Figure 5.4
Figure 5.4. Figure 5.4: Experiment 2: ATE vs. Neuron Importance Band. This line plot shows the TMLE ATE estimate resulting from ablating neurons within specific 5% importance bands (sorted from least to most important based on W1 probe coefficients) for each shared layer. The baseline ATE (…
Figure 5.5
Figure 5.5. Figure 5.5: Decomposed Pathway Quality per Input Neuron (No Effect Data). Stacked bar chart showing sparsity score (blue) and success score (orange) for causal traces originating from each input neuron of the fQ,g model trained on data with no treatment effect [PITH_FULL_IMAGE:…
Figure 5.6
Figure 5.6. Figure 5.6: Pathway Overlap (Jaccard Index) Heatmap (No Effect Data). Shows the overlap between pathways originating from pairs of input neurons for the fQ,g model trained on data with no treatment effect. 5.3 Discovering and Visualizing Treatment vs. Con￾founder Pathways II Usi…
Figure 5.7
Figure 5.7. Figure 5.7: Causal Pathway Visualization (Example: Input Neuron 1). Shows the propagation of effects originating from activating input neuron 1 in the fQ,g model. Edges represent significant activation changes induced by patching. 66 [PITH_FULL_IMAGE:figures/full_fig_p072_5_7.png]
Figure 5.8
Figure 5.8. Figure 5.8: Combined Causal Pathways (High Overlap Example). Overlays pathways from traces initiated from two input neurons with high overlap in the fQ,g model. Green nodes are activated in both traces. 67 [PITH_FULL_IMAGE:figures/full_fig_p073_5_8.png]
Figure 5.9
Figure 5.9. Figure 5.9: Combined Causal Pathways (Low Overlap Example). Overlays pathways from traces initiated from two input neurons with low overlap in the fQ,g model. 68 [PITH_FULL_IMAGE:figures/full_fig_p074_5_9.png]
Figure 1
Figure 1. Figure 1: Training and Validation Losses for ToyTMLE on DS1 (Strong Confounder). 80 [PITH_FULL_IMAGE:figures/full_fig_p086_1.png]
Figure 2
Figure 2. Figure 2: Training and Validation Losses for ToyTMLE on DS2 (No Effect Confounders). Experiment 3: Discovering and Visualizing Treatment vs. Confounder Pathways (Section 5.3) Dataset: DS1 - Strong Confounder • Details: Same as in Experiment 1 (Section 6). Model: fQ,g - Pre-train…
Figure 3
Figure 3. Figure 3: Pathways for confounder input 0 83 [PITH_FULL_IMAGE:figures/full_fig_p089_3.png]
Figure 4
Figure 4. Figure 4: Pathways for confounder input 1 84 [PITH_FULL_IMAGE:figures/full_fig_p090_4.png]
Figure 5
Figure 5. Figure 5: Pathways for confounder input 2 85 [PITH_FULL_IMAGE:figures/full_fig_p091_5.png]
Figure 6
Figure 6. Figure 6: Pathways for confounder input 3 86 [PITH_FULL_IMAGE:figures/full_fig_p092_6.png]
Figure 7
Figure 7. Figure 7: Pathways for confounder input 4 87 [PITH_FULL_IMAGE:figures/full_fig_p093_7.png]
Figure 8
Figure 8. Figure 8: Pathways for confounder input 5 88 [PITH_FULL_IMAGE:figures/full_fig_p094_8.png]
Figure 9
Figure 9. Figure 9: Combined overlap - high interaction between inputs 0 and 2 89 [PITH_FULL_IMAGE:figures/full_fig_p095_9.png]
Figure 10
Figure 10. Figure 10: Combined overlap - low interaction between inputs 2 and 3 90 [PITH_FULL_IMAGE:figures/full_fig_p096_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 43 canonical work pages

  1. [1]

    Jordan, and Tijana Zrnic

    Angelopoulos Anastasiadiou, Stephen Bates, Michael I. Jordan, and Tijana Zrnic. Prediction-powered inference. 2023. URL https://arxiv.org/abs/2301.09633

    arXiv 2023

  2. [2]

    Invariant risk minimization

    Martin Arjovsky, L \'e on Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. URL https://arxiv.org/abs/1907.02893

    arXiv 1907

  3. [3]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. https://arxiv.org/abs/1607.06450, 2016

    arXiv 2016

  4. [4]

    Bang and J

    H. Bang and J. M. Robins. Doubly robust estimation in missing data and causal inference models. Biometrics, 61 0 (4): 0 962--973, 2005

    2005

  5. [5]

    Biopharmaceutical industry-sponsored clinical trials: Impact on state economies, 2015

    Battelle Memorial Institute . Biopharmaceutical industry-sponsored clinical trials: Impact on state economies, 2015. https://orclinicalresearch.com/wp-content/uploads/2018/12/battelle-2015-study.pdf

    work page 2015

  6. [6]

    Probing classifiers: Promises, shortcomings, and advances

    Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48 0 (1): 0 207--219, 2022

    2022

  7. [7]

    Random forests

    Leo Breiman. Random forests. Machine Learning, 45: 0 5--32, 2001

    2001

  8. [8]

    Chan, Adam Jermyn, J

    Trenton Bricken, Collin Templeton, Nicholas Durk, Logan H. Chan, Adam Jermyn, J. D. Chen, Tom Barez, Samuel R. Clayton, Will Gurnee, Cameron Hume, Najib Salim, Tom Henighan, Clark Olsson, Daniel M. Cummings, Brian Hubinger, Cem Anil, Jason Hawkins, Amanda Askell, Nelson Callahan, Robert Huben, Nelson Elhage, and Chris Olah. Sparse autoencoders find highly...

    arXiv 2023

Show all 70 references
  1. [9]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 2020

    2020

  2. [10]

    Causal scrubbing: a method for rigorously testing interpretability hypotheses

    Lawrence Chan, Adrià Garriga-Alonso, Nicholas Goldowsky-Dill, Ryan Greenblatt, Jenny Nitishinskaya, Ansh Radhakrishnan, Buck Shlegeris, and Nate Thomas. Causal scrubbing: a method for rigorously testing interpretability hypotheses. AI Alignment Forum, 2022

    2022

  3. [11]

    D. R. Cox. Regression models and life-tables. Journal of the Royal Statistical Society. Series B (Methodological), 34 0 (2): 0 187--220, 1972

    1972

  4. [12]

    George V. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals and Systems, 2: 0 303--314, 1989

    1989

  5. [13]

    Transcoders find interpretable llm feature circuits

    Jacob Dunefsky, Philippe Chlenski, and Neel Nanda. Transcoders find interpretable llm feature circuits. Advances in Neural Information Processing Systems 37 (NeurIPS 2024), 2024

    2024

  6. [14]

    A mathematical framework for transformer circuits

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dari...

    2021

  7. [15]

    Data on notable ai models, 2024

    Epoch AI . Data on notable ai models, 2024. URL https://epoch.ai/data/notable-ai-models. Accessed: 2025-03-19

    2024

  8. [16]

    Modern robust statistical methods: an easy way to maximize the accuracy and power of your research

    David M Erceg-Hurn and V Mark Mirosevich. Modern robust statistical methods: an easy way to maximize the accuracy and power of your research. Behavior Research Methods, 40 0 (4): 0 898--917, 2008. doi:10.3758/BRM.40.4.898. URL https://pubmed.ncbi.nlm.nih.gov/18855490/

    2008

  9. [17]

    Balance regularized neural network models for causal effect estimation

    Mehrdad Farajtabar, Andrew Lee, Yuanjian Feng, Vishal Gupta, Peter Dolan, and Harish. Balance regularized neural network models for causal effect estimation. arXiv preprint arXiv:2011.11199, 2020. URL https://arxiv.org/abs/2011.11199

    2011 arXiv

  10. [18]

    Schapire

    Yoav Freund and Robert E. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55 0 (1): 0 119--139, August 1997

    1997

  11. [19]

    Friedman

    Jerome H. Friedman. Greedy function approximation: a gradient boosting machine. Annals of Statistics, pages 1189--1232, 2001

    2001

  12. [20]

    Scaling and evaluating sparse autoencoders

    Leo Gao, Tom Dupr \'e la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093, 2024. URL https://arxiv.org/abs/2406.04093. Work from OpenAI

    2024 arXiv

  13. [21]

    Causal abstractions of neural networks

    Atticus Geiger, Zhengxuan Wu, Christopher Potts, Thomas Icard, and Noah D Goodman. Causal abstractions of neural networks. In Advances in Neural Information Processing Systems ( NeurIPS ) , volume 34, pages 10145--10158, 2021. URL https://proceedings.neurips.cc/paper_files/pap...

    2021

  14. [22]

    Inducing causal structure for interpretable neural networks

    Atticus Geiger, Thomas Icard, Christopher Potts, and Noah D Goodman. Inducing causal structure for interpretable neural networks. In Advances in Neural Information Processing Systems ( NeurIPS ) , 2022

    2022

  15. [23]

    Deep Learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, Cambridge, MA, 2016

    2016

  16. [24]

    Natural and Political Observations Mentioned in a Following Index, and Made Upon the Bills of Mortality

    John Graunt. Natural and Political Observations Mentioned in a Following Index, and Made Upon the Bills of Mortality. Printed by Tho. Roycroft for John Martin, James Allestry, and Tho. Dicas, London, 1662. The first edition was published between January 25 and February 5, 1662

  17. [25]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015

    2015 arXiv

  18. [26]

    Burgess, Xavier Glorot, Matthew M

    Irina Higgins, Lo \"i c Matthey, Arka Pal, Christopher P. Burgess, Xavier Glorot, Matthew M. Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representa...

    2017

  19. [27]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840--6851. Curran Associates, Inc., 2020. URL h...

    2020

  20. [28]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9 0 (8): 0 1735--1780, 1997. doi:10.1162/neco.1997.9.8.1735

    1997 doi

  21. [29]

    An artificial neural network for spatio-temporal bipolar patterns: application to phoneme classification

    Toshiteru Homma, Les E Atlas, and Robert J Marks II. An artificial neural network for spatio-temporal bipolar patterns: application to phoneme classification. In Neural information processing systems, pages 31--40, 1987

    1987

  22. [30]

    Multilayer feedforward networks are universal approximators

    Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural Networks, 2 0 (5): 0 359--366, 1989. ISSN 0893-6080. doi:https://doi.org/10.1016/0893-6080(89)90020-8. URL https://www.sciencedirect.com/science/article/pii...

    1989

  23. [31]

    Maximum mean discrepancy for class ratio estimation

    Harish Iyer et al. Maximum mean discrepancy for class ratio estimation. In Proceedings of the 31st International Conference on Machine Learning, volume 32 of PMLR, pages 530--538, 2014. URL https://proceedings.mlr.press/v32/iyer14.html

    2014

  24. [32]

    Johansson, Uri Shalit, and David Sontag

    Fredrik D. Johansson, Uri Shalit, and David Sontag. Learning representations for counterfactual inference. In Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 3020--3029, 2016. URL http://proceed...

    2016

  25. [33]

    Highly accurate protein structure prediction with AlphaFold

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Z \' dek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A A Kohl, Andrew J Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stani...

    2021

  26. [34]

    Linear representations of political perspective emerge in large language models

    Junsol Kim, James Evans, and Aaron Schein. Linear representations of political perspective emerge in large language models. arXiv preprint arXiv:2503.02080, 2025. URL https://arxiv.org/abs/2503.02080

    2025 arXiv

  27. [35]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013

    2013 arXiv

  28. [36]

    Using multiple imputation to deal with missing data and attrition in longitudinal studies with patient-reported outcomes

    Mette Madsen, Bjarne Christensen, Mette T H ybye, Flemming Olesen, and Bodil Nielsen. Using multiple imputation to deal with missing data and attrition in longitudinal studies with patient-reported outcomes. Patient Related Outcome Measures, 6: 0 15--25, 2015. doi:10.2147/PROM...

    2015 doi

  29. [37]

    Auditing language models for hidden objectives, 2025

    Samuel Marks, Johannes Treutlein, Trenton Bricken, Jack Lindsey, Jonathan Marcus, Siddharth Mishra-Sharma, Daniel Ziegler, Emmanuel Ameisen, Joshua Batson, Tim Belonax, Shan Carter, Brian Chen, Hoagy Cunningham, Carson Denison, Florien Dietz, Satvik Golechha, Akbir Khan, Jan K...

    2025

  30. [38]

    Locating and editing factual associations in GPT

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT . In Advances in Neural Information Processing Systems ( NeurIPS ) , 2022. URL https://arxiv.org/abs/2202.05262

    2022 arXiv

  31. [39]

    Meyes, M

    R. Meyes, M. Lu, C. Waubert de Puiseau, and T. Meisen. Ablation studies in artificial neural networks. 2019

    2019

  32. [40]

    Perceptrons: An Introduction to Computational Geometry

    Marvin Minsky and Seymour Papert. Perceptrons: An Introduction to Computational Geometry. MIT Press, 1969

    1969

  33. [41]

    Attribution patching: Activation patching at industrial scale

    Neel Nanda. Attribution patching: Activation patching at industrial scale. https://www.neelnanda.io/mechanistic-interpretability/attribution-patching, Mar 2023. Accessed [Insert Date Accessed]

    2023

  34. [42]

    Niven and H.-Y

    T. Niven and H.-Y. Kao. Probing neural network comprehension of natural language arguments. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4658--4664, 2019

    2019

  35. [43]

    Sparse autoencoders trained on the same data learn different features

    Gonçalo Paulo and Nora Belrose. Sparse autoencoders trained on the same data learn different features. arXiv preprint arXiv:2501.16615, 2025. URL https://arxiv.org/abs/2501.16615

    2025 arXiv

  36. [44]

    Causality: Models, Reasoning, and Inference

    Judea Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, New York, 2nd edition, 2009. ISBN 978-0521895606

    2009

  37. [45]

    Prinja, N

    S. Prinja, N. Gupta, and R. Varma. Censoring in clinical trials: Review of survival analysis techniques. Indian Journal of Community Medicine, 35 0 (2): 0 217--221, 2010

    2010

  38. [46]

    Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders

    Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, Janos Kramar, and Neel Nanda. Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders. arXiv preprint arXiv:2407.14435, 2024. URL https://arxiv.org/abs/2407.14435

    2024 arXiv

  39. [47]

    Stochastic backpropagation and approximate inference in deep generative models

    Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In Proceedings of the International Conference on Machine Learning, 2014

    2014

  40. [48]

    The perceptron: A probabilistic model for information storage and organization in the brain

    Frank Rosenblatt. The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65 0 (6): 0 386--408, 1958. doi:10.1037/h0042519

    1958 doi

  41. [49]

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internal representations by error propagation. Technical Report ICS 8504, Institute for Cognitive Science, University of California, San Diego, 1985

    1985

  42. [50]

    Polysemanticity and capacity in neural networks

    Adam Scherlis, Kshitij Mueller, János András Dang, Nelson Elhage, and Chris Olah. Polysemanticity and capacity in neural networks. arXiv preprint arXiv:2210.01892, 2022. URL https://arxiv.org/abs/2210.01892

    2022 arXiv

  43. [51]

    Toward causal representation learning

    Bernhard Sch \"o lkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE, 109: 0 612--634, 2021

    2021

  44. [52]

    Adjustment for confounding using pre-trained representations

    Rickmer Schulte, David R \"u gamer, and Thomas Nagler. Adjustment for confounding using pre-trained representations. In ICLR 2025 Workshop FM-Wild, March 2025. Submission Number: 98

    2025

  45. [53]

    271 the neyman— rubin model of causal inference and estimation via matching methods

    Jasjeet Sekhon. 271 the neyman— rubin model of causal inference and estimation via matching methods. In The Oxford Handbook of Political Methodology. Oxford University Press, 08 2008. ISBN 9780199286546. doi:10.1093/oxfordhb/9780199286546.003.0011. URL https://doi.org/10.1093/...

    2008

  46. [54]

    Johansson, and David Sontag

    Uri Shalit, Fredrik D. Johansson, and David Sontag. Estimating individual treatment effect: generalization bounds and algorithms. Proceedings of the 34th International Conference on Machine Learning, 70: 0 3073--3082, 2017

    2017

  47. [55]

    Shi, Victor Veitch, and David M

    C. Shi, Victor Veitch, and David M. Blei. Invariant representation learning for treatment effect estimation. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, volume 161 of Proceedings of Machine Learning Research, pages 1546--1555. PMLR, 2021

    2021

  48. [56]

    Blei, and Victor Veitch

    Claudia Shi, David M. Blei, and Victor Veitch. Adapting neural networks for the estimation of treatment effects. arXiv preprint, 2019 a . URL https://www.cs.columbia.edu/ blei/papers/ShiBleiVeitch2019.pdf

    2019

  49. [57]

    Blei, and Victor Veitch

    Claudia Shi, David M. Blei, and Victor Veitch. Adapting neural networks for the estimation of treatment effects. In Advances in Neural Information Processing Systems 32 (NeurIPS 2019), Vancouver, Canada, December 2019 b

    2019

  50. [58]

    Longitudinal targeted minimum loss-based estimation with temporal-difference heterogeneous transformer

    Toru Shirakawa, Yi Li, Yulun Wu, Sky Qiu, Yuxuan Li, Mingduo Zhao, Hiroyasu Iso, and Mark van der Laan. Longitudinal targeted minimum loss-based estimation with temporal-difference heterogeneous transformer. International Conference of Machine Learning, 2024

    2024

  51. [59]

    To explain or to predict? Statistical Science, 25 0 (3): 0 289--310, 2010

    Galit Shmueli. To explain or to predict? Statistical Science, 25 0 (3): 0 289--310, 2010. doi:10.1214/10-STS330

    2010 doi

  52. [60]

    Regression shrinkage and selection via the Lasso

    Robert Tibshirani. Regression shrinkage and selection via the Lasso . Journal of the Royal Statistical Society. Series B (Methodological), 58 0 (1): 0 267--288, 1996

    1996

  53. [61]

    van der Laan and James M

    Mark J. van der Laan and James M. Robins. Unified methods for censored longitudinal data and causality. Springer, 2003

    2003

  54. [62]

    van der Laan and Sherri Rose

    Mark J. van der Laan and Sherri Rose. Targeted Learning: Causal Inference for Observational and Experimental Data. Springer New York, 2013. ISBN 9781461429111

    2013

  55. [64]

    van der Laan and Daniel Rubin

    Mark J. van der Laan and Daniel Rubin. Targeted maximum likelihood learning. U.C. Berkeley Division of Biostatistics Working Paper Series, 2006

    2006

  56. [65]

    van der Laan, Eric C

    Mark J. van der Laan, Eric C. Polley, and Alan E. Hubbard. Super learner. Statistical Applications in Genetics and Molecular Biology, 6 0 (1), 2007. doi:10.2202/1544-6115.1309

    2007

  57. [66]

    van der Laan, Maya L

    Mark J. van der Laan, Maya L. Petersen, and Alan E. Hubbard. Conducting strategic and scientific research using observational data. Journal of the National Cancer Institute, 100 0 (5): 0 332--335, 2008. doi:10.1093/jnci/djn032

    2008 doi

  58. [68]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, pa...

    2017

  59. [69]

    Deep learning-based propensity scores for confounding control in comparative effectiveness research: A large-scale, real-world data study

    Janick Weberpals, Tim Becker, Jessica Davies, Fabian Schmich, Dominik R \"u ttinger, Fabian J Theis, and Anna Bauer-Mehren. Deep learning-based propensity scores for confounding control in comparative effectiveness research: A large-scale, real-world data study. Epidemiology, ...

    2021 doi

  60. [70]

    Causal proxy models for concept-based model explanations

    Zhengxuan Wu, Karel D'Oosterlinck, Atticus Geiger, Amir Zur, and Christopher Potts. Causal proxy models for concept-based model explanations. arXiv preprint arXiv:2209.14279, September 2022

    2022 arXiv

  61. [71]

    van der Laan

    Wenjing Zheng and Mark J. van der Laan. Mediation Analysis with Time-Varying Mediators and Exposures, pages 277--299. Springer International Publishing, Cham, 2018. ISBN 978-3-319-65304-4. doi:10.1007/978-3-319-65304-4_10. URL https://doi.org/10.1007/978-3-319-65304-4_10

    2018 doi

  62. [72]

    Ensemble Methods: Foundations and Algorithms

    Zhi-Hua Zhou. Ensemble Methods: Foundations and Algorithms. CRC Press, Boca Raton, FL, 2012. ISBN 9781439830031

    2012

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.