pith. sign in

arxiv: 2505.03530 · v3 · submitted 2025-05-06 · 💻 cs.LG

A Multi-Level Causal Intervention Framework for Mechanistic Interpretability in Variational Autoencoders

Dip Roy , Rajiv Misra , Sanjay Kumar Singh , Anisha Roy This is my paper

Pith reviewed 2026-05-22 16:33 UTC · model grok-4.3

classification 💻 cs.LG
keywords variational autoencodersmechanistic interpretabilitycausal interventionsdisentanglementgenerative modelsactivation patchingcausal mediation
0
0 comments X

The pith

A multilevel causal intervention framework interprets variational autoencoders and reveals architecture trade-offs.

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

This paper develops a general-purpose framework for probing the internal mechanisms of variational autoencoders using causal interventions at input, latent, and activation levels. It introduces new metrics to measure causal effect strength, intervention specificity, and circuit modularity that go beyond traditional disentanglement measures. Experiments across six VAE variants and five datasets reveal a consistent trade-off where higher causal effect strength correlates with lower disentanglement scores. The study also shows that architecture performance depends on dataset structure and that standard metrics break down for discrete latent representations.

Core claim

The paper establishes a multilevel causal intervention framework for VAEs that includes input manipulation, latent-space perturbation, activation patching, and causal mediation analysis, along with three new metrics: Causal Effect Strength, intervention specificity, and circuit modularity. This framework is used to conduct a large-scale empirical study showing a negative correlation between CES and DCI disentanglement, capacity bottlenecks in beta-VAE on complex data, no universally best architecture, and limitations of continuous metrics on discrete spaces.

What carries the argument

The multilevel causal intervention framework consisting of four manipulation types and three new quantitative metrics for assessing causal properties in VAE representations.

Where Pith is reading between the lines

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

  • The framework could be adapted to compare causal structures in other generative models like diffusion models.
  • Practitioners might use the new metrics to select or regularize VAE architectures based on dataset complexity.
  • Extending the approach to larger models could test whether the CES-DCI trade-off scales with model size.

Load-bearing premise

The defined interventions isolate causal mechanisms in VAEs without confounding effects from training or architecture.

What would settle it

If targeted causal mediation on one latent factor fails to change only the corresponding output features while leaving others unaffected, the framework's isolation of mechanisms would be undermined.

Figures

Figures reproduced from arXiv: 2505.03530 by Anisha Roy, Dip Roy, Rajiv Misra, Sanjay Kumar Singh.

Figure 11
Figure 11. Figure 11: Training loss curves on dSprites for all six architectures across 30 epochs. All continuous VAE variants [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 1
Figure 1. Figure 1: Latent traversals for β-VAE on 3DShapes. Each row corresponds to a single latent dimension swept from −3 to +3. Individual rows control distinct factors (object hue, wall hue, floor hue, scale, orientation), demonstrating disentangled encoding (DCI = 0.805) [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Latent traversals for Standard VAE on 3DShapes. Multiple factors change simultaneously within single [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-dimension Causal Effect Strength (left) and Intervention Specificity (right) on dSprites. β-VAE (orange) shows reduced CES across most dimensions compared to Standard VAE (green), FactorVAE (pink), and β-TC-VAE (teal). VQ-VAE (yellow) shows near-zero CES but anomalously high specificity due to the discrete codebook [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Causal mediation heatmap for β-VAE on dSprites. Rows represent generative factors; columns represent encoder layers. Mediation strength is concentrated in encoder_conv_0 and encoder_conv_1 (values ~0.12–0.15), dropping sharply at encoder_conv_2 (~0.001). This indicates that factor-specific information is primarily processed in early convolutional layers [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 8
Figure 8. Figure 8: CES–DCI relationship analysis. (a) Scatter plot of average CES vs. DCI disentanglement across all datasets and continuous VAE architectures, showing a positive cross-dataset trend (r = 0.534, p = 0.018) driven by dataset complexity differences. (b) Within-dataset Pearson correlations reveal strong negative correlations on dSprites (r = −0.95, p = 0.014) and 3DShapes (r = −0.98, p = 0.023), confirming the C… view at source ↗
Figure 9
Figure 9. Figure 9: KL divergence per latent dimension on dSprites for all six architectures. β [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 5
Figure 5. Figure 5: Circuit modularity across network layers on 3DShapes. Modularity is concentrated in the mu (latent) layer, [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Disentanglement metrics comparison on 3DShapes across all six VAE architectures. β [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 10
Figure 10. Figure 10: presents the cross-dataset summary comparing all architectures simultaneously, providing a visual overview of how disentanglement varies with dataset complexity [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 7
Figure 7. Figure 7: Hyperparameter sweep on dSprites. (a) β-VAE shows a clear CES–DCI tradeoff as β increases: DCI rises from 0.094 to 0.573 while CES drops from 8.1 to 3.1. (b) FactorVAE is completely insensitive to γ, with DCI and CES essentially unchanged across an 8× range. (c) Combined view showing the two distinct behavioral families in CES–DCI space [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 12
Figure 12. Figure 12: Quantitative ablation for FactorVAE on CelebA. Panel (a) shows raw metric values under each ablation [PITH_FULL_IMAGE:figures/full_fig_p027_12.png] view at source ↗
read the original abstract

Understanding how generative models represent and transform data is a foundational problem in deep learning interpretability. While mechanistic interpretability of discriminative architectures has yielded substantial insights, relatively little work has addressed variational autoencoders (VAEs). This paper presents the first general-purpose multilevel causal intervention framework for mechanistic interpretability of VAEs. The framework comprises four manipulation types: input manipulation, latent-space perturbation, activation patching, and causal mediation analysis. We also define three new quantitative metrics capturing properties not measured by existing disentanglement metrics alone: Causal Effect Strength (CES), intervention specificity, and circuit modularity. We conduct the largest empirical study to date of VAE causal mechanisms across six architectures (standard VAE, beta-VAE, FactorVAE, beta-TC-VAE, DIP-VAE-II, and VQ-VAE) and five benchmarks (dSprites, 3DShapes, MPI3D, CelebA, and SmallNORB), with three seeds per configuration, totaling 90 independent training runs. Our results reveal several findings: (i) a consistent within-dataset negative correlation between CES and DCI disentanglement (the CES-DCI trade-off); (ii) that the KL reweighting mechanism of beta-VAE induces a capacity bottleneck when generative factors approach latent dimensionality, degrading disentanglement on complex datasets; (iii) that no single VAE architecture dominates across all five datasets, with optimal choice depending on dataset structure; and (iv) that CES-based metrics applied to discrete latent spaces (VQ-VAE) yield near-zero values, revealing a critical limitation of continuous-intervention methods for discrete representations. These results provide both a theoretical foundation and comprehensive empirical evaluation for mechanistic interpretability of generative models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents the first general-purpose multilevel causal intervention framework for mechanistic interpretability of VAEs. The framework includes four manipulation types (input manipulation, latent-space perturbation, activation patching, and causal mediation analysis) and defines three new metrics: Causal Effect Strength (CES), intervention specificity, and circuit modularity. It reports results from the largest empirical study to date, covering six VAE architectures and five datasets across 90 independent training runs, including findings on a CES-DCI trade-off, effects of KL reweighting in beta-VAE, architecture-dependent performance, and limitations of continuous interventions on discrete latents like VQ-VAE.

Significance. If the interventions validly isolate causal mechanisms, the work would provide a useful foundation and metrics for mechanistic interpretability of generative models, extending beyond existing disentanglement measures. The scale of the study (six architectures, five benchmarks, three seeds) and explicit reporting of multiple configurations are strengths that support reproducible comparisons and the observed trade-offs. The empirical findings on architecture-dataset interactions and discrete latent limitations add practical value.

major comments (2)
  1. [Section 3] The core claim that the four interventions cleanly reveal causal structure rests on the untested assumption that they are unconfounded by joint encoder-decoder training and ELBO optimization. In the intervention definitions and implementation details (Section 3), the manuscript should include ablations or controls demonstrating that CES and related metrics are not driven by reconstruction biases or variational posterior artifacts; without this, the CES-DCI trade-off and architecture comparisons lose mechanistic grounding.
  2. [Results section] The reported negative CES-DCI correlation is load-bearing for the trade-off claim. The results section should report statistical significance (e.g., p-values or confidence intervals) and controls for dataset complexity or latent dimensionality to confirm the correlation is not an artifact of specific benchmarks like dSprites versus CelebA.
minor comments (2)
  1. [Abstract] The abstract states 'the largest empirical study to date' but should explicitly note the total of 90 runs for immediate clarity.
  2. [Methods] Ensure the exact formulas for intervention specificity and circuit modularity are presented with consistent notation and pseudocode in the methods to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. We address each major comment below and have incorporated revisions to strengthen the empirical grounding of our claims.

read point-by-point responses

  1. Referee: [Section 3] The core claim that the four interventions cleanly reveal causal structure rests on the untested assumption that they are unconfounded by joint encoder-decoder training and ELBO optimization. In the intervention definitions and implementation details (Section 3), the manuscript should include ablations or controls demonstrating that CES and related metrics are not driven by reconstruction biases or variational posterior artifacts; without this, the CES-DCI trade-off and architecture comparisons lose mechanistic grounding.

    Authors: We appreciate the referee's emphasis on rigorously ruling out confounds from joint training and ELBO optimization. Our intervention definitions aim to isolate effects at distinct levels, but we agree that explicit ablations would provide stronger evidence that CES, specificity, and modularity reflect causal mechanisms rather than reconstruction biases or posterior artifacts. In the revised manuscript, we will expand Section 3 with new controls, including (i) interventions on models with separately trained encoders/decoders and (ii) comparisons against fixed variational posteriors, to demonstrate that the metrics remain stable and retain their interpretive value. These additions will directly support the validity of the CES-DCI trade-off and cross-architecture comparisons. revision: yes

  2. Referee: [Results section] The reported negative CES-DCI correlation is load-bearing for the trade-off claim. The results section should report statistical significance (e.g., p-values or confidence intervals) and controls for dataset complexity or latent dimensionality to confirm the correlation is not an artifact of specific benchmarks like dSprites versus CelebA.

    Authors: We agree that statistical rigor and explicit controls are necessary to substantiate the CES-DCI trade-off. In the revised results section, we will report p-values and 95% confidence intervals for all within-dataset correlations. To address potential artifacts from benchmark-specific factors, we will include additional analyses such as partial correlations controlling for dataset complexity (measured by number of generative factors and image resolution) and latent dimensionality. While our study already spans five datasets with varying complexities and six architectures with different latent sizes, these controls will confirm that the negative correlation is not driven by particular dataset-architecture pairings. revision: yes

Circularity Check

0 steps flagged

No circularity: framework definitions and empirical observations are independent of inputs

full rationale

The paper defines a multilevel causal intervention framework consisting of input manipulation, latent perturbation, activation patching, and causal mediation analysis, then introduces three new metrics (CES, intervention specificity, circuit modularity) as direct functions of those intervention outcomes on trained VAEs. The reported findings, including the CES-DCI trade-off, architecture comparisons, and limitations for discrete latents, are presented as results from 90 independent empirical runs across datasets and models rather than any first-principles derivation or prediction that reduces to the definitions or fitted parameters by construction. No self-citation load-bearing steps, uniqueness theorems, or ansatzes are invoked in the provided text; the contribution is self-contained as definitional plus observational.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The central claim rests on domain assumptions about causality in neural networks and introduces several new metrics and concepts without external independent evidence or formal proofs.

axioms (1)
  • domain assumption Interventions at input, latent, and activation levels correspond to causal manipulations in the generative process of VAEs.
    This underpins the validity of the four manipulation types and the causal mediation analysis.
invented entities (3)
  • Causal Effect Strength (CES) no independent evidence
    purpose: Quantify the magnitude of causal effects from interventions in VAE latent spaces.
    Newly defined metric without reference to prior independent validation or theoretical derivation.
  • intervention specificity no independent evidence
    purpose: Measure how targeted an intervention is to specific generative factors.
    Introduced as part of the new quantitative metrics.
  • circuit modularity no independent evidence
    purpose: Assess the modularity of computational circuits within the VAE.
    New metric for properties not captured by disentanglement alone.

pith-pipeline@v0.9.0 · 5850 in / 1411 out tokens · 113703 ms · 2026-05-22T16:33:37.890657+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Why Do Large Language Models Generate Harmful Content?

    cs.AI 2026-04 unverdicted novelty 6.0

    Causal mediation analysis shows harmful LLM outputs arise in late layers from MLP failures and gating neurons, with early layers handling harm context detection and signal propagation.

  2. Posterior-Calibrated Causal Circuits in Variational Autoencoders: Why Image-Domain Interpretability Fails on Tabular Data

    cs.LG 2026-03 unverdicted novelty 6.0

    Tabular VAEs show ~50% lower causal circuit modularity than image VAEs, with beta-VAE CES collapsing to 0.043 versus 0.133 due to reconstruction degradation, challenging direct transfer of image interpretability techniques.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · cited by 2 Pith papers

  1. [1]

    Feature visualization,

    C. Olah, A. Mordvintsev, and L. Schubert, "Feature visualization," Distill, vol. 2, no. 11, p. e7, 2017

    work page 2017

  2. [2]

    A mathematical framework for transformer circuits,

    N. Elhage et al., "A mathematical framework for transformer circuits," Anthropic, Tech. Rep., 2021

    work page 2021

  3. [3]

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

    K. Wang et al., "Interpretability in the wild: A circuit for indirect object identification in GPT -2 small," in Proc. Int. Conf. Learn. Represent. (ICLR), 2023

    work page 2023

  4. [4]

    Network dissection: Quantifying interpretability of deep visual representations,

    D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba, "Network dissection: Quantifying interpretability of deep visual representations," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 6541 –6549

    work page 2017

  5. [5]

    Auto-encoding variational Bayes,

    D. P. Kingma and M. Welling, "Auto-encoding variational Bayes," in Proc. Int. Conf. Learn. Represent. (ICLR), 2014

    work page 2014

  6. [6]

    β-VAE: Learning basic visual concepts with a constrained variational framework,

    I. Higgins et al., "β-VAE: Learning basic visual concepts with a constrained variational framework," in Proc. Int. Conf. Learn. Represent. (ICLR), 2017

    work page 2017

  7. [7]

    Disentangling by factorising,

    H. Kim and A. Mnih, "Disentangling by factorising," in Proc. Int. Conf. Mach. Learn. (ICML), 2018, pp. 2649 – 2658

    work page 2018

  8. [8]

    Isolating sources of disentanglement in variational autoencoders,

    R. T. Q. Chen, X. Li, R. Grosse, and D. Duvenaud, "Isolating sources of disentanglement in variational autoencoders," in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 31, 2018

    work page 2018

  9. [9]

    Variational inference of disentangled latent concepts from unlabeled observations,

    A. Kumar, P. Sattigeri, and A. Balakrishnan, "Variational inference of disentangled latent concepts from unlabeled observations," in Proc. Int. Conf. Learn. Represent. (ICLR), 2018

    work page 2018

  10. [10]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, "Neural discrete representation learning," in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 30, 2017

    work page 2017

  11. [11]

    A framework for the quantitative evaluation of disentangled representations,

    C. Eastwood and C. K. I. Williams, "A framework for the quantitative evaluation of disentangled representations," in Proc. Int. Conf. Learn. Represent. (ICLR), 2018

    work page 2018

  12. [12]

    Separated attribute predictability (SAP) score,

    A. Kumar, P. Sattigeri, and A. Balakrishnan, "Separated attribute predictability (SAP) score," in Workshop Adv. Neural Inf. Process. Syst., 2018

    work page 2018

  13. [13]

    Progress measures for grokking via mechanistic interpretability,

    N. Nanda, L. Chan, T. Lieberum, J. Smith, and J. Steinhardt, "Progress measures for grokking via mechanistic interpretability," in Proc. Int. Conf. Learn. Represent. (ICLR), 2023

    work page 2023

  14. [14]

    Challenging common assumptions in the unsupervised learning of disentangled representations,

    F. Locatello et al., "Challenging common assumptions in the unsupervised learning of disentangled representations," in Proc. Int. Conf. Mach. Learn. (ICML), 2019, pp. 4114 –4124

    work page 2019

  15. [15]

    Causal abstractions of neural networks,

    A. Geiger, H. Lu, T. Icard, and C. Potts, "Causal abstractions of neural networks," in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 34, 2021

    work page 2021

  16. [16]

    Locating and editing factual associations in GPT,

    K. Meng, D. Bau, A. Andonian, and Y. Belinkov, "Locating and editing factual associations in GPT," in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 35, 2022

    work page 2022

  17. [17]

    Investigating gender bias in language models using causal mediation analysis,

    J. Vig, S. Gehrmann, Y. Belinkov, S. Qian, D. Nishi, Y. Zhang, and Y. Jernite, "Investigating gender bias in language models using causal mediation analysis," in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020

    work page 2020

  18. [18]

    Curve circuits,

    N. Cammarata, S. Carter, G. Goh, C. Olah, M. Petrov, and L. Schubert, "Curve circuits," Distill, 2021

    work page 2021

  19. [19]

    dSprites: Disentanglement testing sprites dataset,

    L. Matthey, I. Higgins, D. Hassabis, and A. Lerchner, "dSprites: Disentanglement testing sprites dataset," GitHub Repository, 2017

    work page 2017

  20. [20]

    3D shapes dataset,

    C. Burgess and H. Kim, "3D shapes dataset," GitHub Repository, 2018

    work page 2018

  21. [21]

    On the transfer of inductive bias from simulation to the real world: A new disentanglement dataset,

    M. Gondal et al., "On the transfer of inductive bias from simulation to the real world: A new disentanglement dataset," in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 32, 2019

    work page 2019

  22. [22]

    Deep learning face attributes in the wild,

    Z. Liu, P. Luo, X. Wang, and X. Tang, "Deep learning face attributes in the wild," in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), 2015, pp. 3730–3738. Causal Intervention Framework for VAE Mechanistic Interpretability 33

    work page 2015

  23. [23]

    Learning methods for generic object recognition with invariance to pose and lighting,

    Y. LeCun, F. J. Huang, and L. Bottou, "Learning methods for generic object recognition with invariance to pose and lighting," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), vol. 2, 2004, pp. II –97

    work page 2004

  24. [24]

    Experiment tracking with Weights and Biases,

    L. Biewald, "Experiment tracking with Weights and Biases," Software available from wandb.com, 2020

    work page 2020

  25. [25]

    Deep learning and the information bottleneck principle,

    N. Tishby and N. Zaslavsky, "Deep learning and the information bottleneck principle," in Proc. IEEE Inf. Theory Workshop (ITW), 2015, pp. 1–5

    work page 2015

  26. [26]

    Pearl, Causality: Models, Reasoning, and Inference, 2nd ed

    J. Pearl, Causality: Models, Reasoning, and Inference, 2nd ed. Cambridge, U.K.: Cambridge Univ. Press, 2009

    work page 2009

  27. [27]

    InfoVAE: Balancing learning and inference in variational autoencoders,

    S. Zhao, J. Song, and S. Ermon, "InfoVAE: Balancing learning and inference in variational autoencoders," in Proc. AAAI Conf. Artif. Intell., vol. 33, 2019, pp. 5885–5892

    work page 2019

  28. [28]

    Theory and evaluation metrics for learning disentangled representations,

    K. Do and T. Tran, "Theory and evaluation metrics for learning disentangled representations," in Proc. Int. Conf. Learn. Represent. (ICLR), 2020

    work page 2020

  29. [29]

    Visualizing and understanding generative adversarial networks,

    D. Bau, J.-Y. Zhu, H. Strobelt, A. Lapedriza, B. Zhou, and A. Torralba, "Visualizing and understanding generative adversarial networks," in Proc. Int. Conf. Learn. Represent. (ICLR), 2019

    work page 2019

  30. [30]

    Testing relational understanding in text-guided image generation,

    C. Conwell, D. Mayo, M. Barbu, G. Buice, M. Cusimano, and B. Katz, "Testing relational understanding in text-guided image generation," arXiv preprint arXiv:2208.00005, 2022

    work page arXiv 2022

  31. [31]

    CausalVAE: Disentangled representation learning via neural structural causal models,

    M. Yang, F. Liu, Z. Chen, X. Shen, J. Hao, and J. Wang, "CausalVAE: Disentangled representation learning via neural structural causal models," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 9593 – 9602

    work page 2021

  32. [32]

    Robustly disentangled causal mechanisms: Validating deep representations for interventional robustness,

    R. Suter, D. Miladinovic, B. Schölkopf, and S. Bauer, "Robustly disentangled causal mechanisms: Validating deep representations for interventional robustness," in Proc. Int. Conf. Mach. Learn. (ICML), 2019, pp. 6056 –6065

    work page 2019

  33. [33]

    Disentangling disentanglement in variational autoencoders,

    E. Mathieu, T. Rainforth, N. Siddharth, and Y. W. Teh, "Disentangling disentanglement in variational autoencoders," in Proc. Int. Conf. Mach. Learn. (ICML), 2019, pp. 4402–4412

    work page 2019

  34. [34]

    Similarity of neural network representations revisited,

    S. Kornblith, M. Norouzi, H. Lee, and G. Hinton, "Similarity of neural network representations revisited," in Proc. Int. Conf. Mach. Learn. (ICML), 2019, pp. 3519–3529

    work page 2019

  35. [35]

    Deep variational information bottleneck,

    A. A. Alemi, I. Fischer, J. V. Dillon, and K. Murphy, "Deep variational information bottleneck," in Proc. Int. Conf. Learn. Represent. (ICLR), 2017

    work page 2017