Pith. sign in

REVIEW 3 major objections 5 minor 18 references

SVI-DAG: A Structured Variational Inference Approach to Bayesian Causal Discovery

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SVI-DAG couples flows and Stein particles to learn calibrated DAG posteriors.

desk verdict SVI-DAG combines normalizing flows over edge logits with SVGD on order potentials—a genuinely new mix—but the abstract's UQ claims are undercut by the paper's own caveats and the mode-coverage mechanism uses a straight-through score that the paper admits is inexact. read the letter →

arxiv 2608.04930 v1 pith:BCGPVSMT submitted 2026-08-05 cs.LG cs.AI

Shrenik Zinage This is my paper
classification cs.LGcs.AI MSC 62F1562D20
keywords BayesiancausaldiscoverystructuredvariationalinferencenormalizingflowsSteingradientdescentDAGposteriorepistemicuncertaintydomain-informedprioracyclicitybyconstruction
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

The paper proposes SVI-DAG, a Bayesian method for causal discovery that aims to approximate the posterior distribution over directed acyclic graphs rather than returning a single graph. Its central claim is that coupling a normalizing-flow guide over edge logits with Stein variational gradient descent over order-potential particles in a relaxed acyclicity space yields a multimodal DAG posterior that is better calibrated than five Bayesian baselines while staying structurally competitive. The authors argue this matters because calibrated epistemic uncertainty over causal structures is what lets practitioners reason about interventions when data are scarce or the DAG is identifiable only up to its Markov equivalence class, and because the method gives a principled way to fold domain knowledge into the search through a Beta-Bernoulli prior. If the claim holds, SVI-DAG is a flow-agnostic recipe for Bayesian causal discovery with explicit mode coverage.

What carries the argument

The load-bearing object is the structural guide pair $(\gamma, r)$ together with the acyclicity-by-construction adjacency $A = B \odot M(r)$, where $M(r) = P(r) L P(r)^T$ is the order mask induced by sorting node potentials $r$ and $L$ is a fixed strictly lower-triangular matrix. The flow $T_\phi(z;r)$ makes the edge-logit distribution conditional and multimodal-capable; the SVGD update over particles $\{r^{(k)}\}$ uses a pullback kernel $k_r(r, r') = k_M(\psi(r), \psi(r'))$ with $\psi(r) = \operatorname{vec}(M_{\tau,KS}(r))$, so the repulsive term acts in relaxed acyclicity space rather than raw potential space. The Beta-Bernoulli prior on edge probabilities, with $\alpha = \nu p + 1$ and $\beta = \nu(1-p) + 1$ and concentration $\nu$ growing with $|p - 0.5|$, induces a Logistic-Beta density on $\gamma_{ij} = \operatorname{logit}(\pi_{ij})$, giving a proper continuous prior that makes the reverse-KL objective finite and carries domain knowledge into the search.

What would settle it

On a synthetic posterior with two well-separated DAG modes that imply different topological orders, compare SVI-DAG's particle diversity and Brier score under the straight-through score used in Algorithm 1, line 23, against the same algorithm using an unbiased estimator of the hard-threshold score; if the two versions differ substantially, the mode-coverage claim is carried by the surrogate, not by the SVGD construction.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a structured variational inference algorithm whose guide no longer treats edges as independent. A conditional normalizing flow $T_\phi(z;r)$ turns base noise into edge logits given node potentials $r$, so dependencies among edges are modeled explicitly; acyclicity is enforced by construction through $A = B \odot M(r)$, where $M(r)$ sorts nodes by $r$ and keeps only lower-triangular edges; and the node potentials are updated by SVGD with a kernel defined on relaxed acyclicity masks, adding repulsive pressure intended to keep multiple posterior modes populated. The paper also introduces a hierarchical Beta-Bernoulli prior over edge probabilities whose hyperparameters concentrate more for beliefs far from $0.5$, which after marginalization induces a Logistic-Beta prior on edge logits that makes the KL divergence well defined and lets prior beliefs act as an inductive bias. Empirically the paper reports that SVI-DAG achieves lower Brier scores and higher AUROC than ProDAG, BayesDAG, DDS, BCD Nets, and DiBS on most linear and nonlinear synthetic settings while remaining competitive in expected structural Hamming distance and F1, and that it leads on F1 for DAG metrics and AUROC for CPDAG metrics on the Sachs data.

Load-bearing premise

The promised mode-covering diversity rests on replacing an exact score gradient with a straight-through surrogate through a hard threshold; if that surrogate is far from the true gradient, the repulsive particle behavior is a heuristic rather than an established property.

Editorial extensions

If this is right

  • Posterior edge probabilities from SVI-DAG give lower Brier scores and higher AUROC than ProDAG, BayesDAG, DiBS, DDS, and BCD Nets on most linear and nonlinear synthetic settings, while expected SHD and F1 stay competitive.
  • The Beta-Bernoulli prior lets domain knowledge act as an inductive bias: on a non-identifiable two-node problem, a strong incorrect prior pushes 84.7% of posterior mass to the wrong orientation, while a strong correct prior pushes 83.2% to the true one.
  • Because every sampled graph $A = B \odot M(r)$ is a hard binary DAG by construction, posterior samples can be used directly for downstream causal queries without rounding or projection.
  • The formulation is flow-agnostic; the reported numbers use neural spline flows, but the algorithm accepts any normalizing flow in place of $T_\phi$.
  • In the identifiable nonlinear setting, the prior acts mainly as a regularizer: an incorrect prior suppresses both edges, concentrating 86.4% of mass on the empty graph, while a correct prior sharpens the posterior to 95.2% on the true orientation.

Reading between the lines

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

  • Because the ELBO is optimized in $(B, r)$ space rather than the tighter $A$ space, part of the reported calibration gain may come from the prior and guide structure rather than from a tighter bound; a direct comparison against an $A$-space bound would isolate this.
  • The SVGD kernel defined on relaxed acyclicity masks is order-based and could be dropped into other order-based DAG samplers, potentially improving their mode coverage as well.
  • A natural stress test: vary the flow family (affine versus spline) and check whether the Brier/AUROC ranking against baselines is stable; the paper's caveat that all results use one flow family suggests this ranking may shift.
  • On non-identifiable linear problems, a fully calibrated posterior should become uniform over the Markov equivalence class as sample size grows; checking this directly would be a stronger calibration test than Brier score alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes SVI-DAG, a structured variational inference method for Bayesian causal discovery. It places a hierarchical Beta-Bernoulli prior over edge indicators with concentration depending on prior strength, maps edge logits through a conditional normalizing flow, enforces acyclicity via a permutation mask built from node-order potentials, and updates those potentials with Stein variational gradient descent using a kernel in a relaxed acyclicity space. The authors report experiments on linear and nonlinear synthetic data and on the Sachs data set, and claim superior uncertainty quantification and competitive structural accuracy against five Bayesian baselines. The code is provided.

Significance. The Beta-Bernoulli marginalization, the induced Logistic-Beta prior, the closed-form Gaussian KL in Eq. (4), and the KL data-processing inequality in Proposition B.2 are correct and clearly presented. The idea of decoupling edge dependencies via flows and order-potential particles via SVGD is interesting, and the release of code is a practical strength. However, the paper's two load-bearing pillars—the exactness of the SVGD mode-coverage mechanism and the empirical superiority claim—are both undermined by the authors' own statements, so the significance as currently evidenced is limited.

major comments (3)
  1. [Section 3.6, Algorithm 1 line 23] Equation (6) defines the SVGD update using the exact score \nabla_r log p~_phi(r|D) = \nabla_r[L(phi;r) + log p(r)], where L(phi;r) in Eq. (5) contains the expectation over the hard threshold B = 1{gamma + R > 0}. The implementation in Algorithm 1 line 23 replaces this with a straight-through gradient through that threshold. Section 3.6 states: 'The derivative of a sampled hard threshold is not this exact derivative. It would require an analytic finite sum or an unbiased discrete score estimator.' The hard-threshold function is piecewise constant, so differentiating through the sampled value is a surrogate, not the score of the stated target. Consequently, the repulsive term in Eq. (6) is not a standard SVGD update for p~_phi, and the claimed mode-coverage and repulsive-diversity properties are heuristic. Section F.0.2 further acknowledges that the method 'does not guarantee recovery of all DAGs within MEC.' The paper should either implement an unbiased score estimator or explicitly reframe the algorithm as optimizing a surrogate target and provide independent evidence that the surrogate preserves diversity.
  2. [Section 4.4, Tables 1 and 2] Section 4.4 contains the statement that all SVI-DAG results use a single flow family, that performance is 'largely governed by this choice,' and that 'the comparisons presented here should not be taken too seriously.' This directly disclaims the abstract's claim of 'superior performance in uncertainty quantification.' Moreover, in Tables 1 and 2 SVI-DAG is not uniformly best: on the Sachs DAG metrics, ProDAG and BCD Nets report lower Brier scores and lower expected SHD, and DiBS reports higher AUROC; on the CPDAG metrics, DiBS reports a lower Brier score and a slightly higher expected F1. Without a principled selection of the flow family and a benchmark protocol the authors are willing to stand behind, the empirical support for the central claim is absent.
  3. [Appendix E, nonlinear synthetic benchmarks] Appendix E states that BCD Nets and DDS 'used the same linear structural model for the linear ER, nonlinear ER, and Sachs experiments' and that 'neither wrapper selected a separate nonlinear architecture for the nonlinear ER benchmarks.' The nonlinear synthetic comparisons in Figures 6 and 7 therefore handicap these two baselines by forcing a linear mechanism on nonlinear data. Since the paper claims to remain 'competitive in terms of structural accuracy' in the nonlinear setting, these results do not provide a fair test of that claim.
minor comments (5)
  1. [Section 5] The word 'quantifcation' in the conclusion should be 'quantification.'
  2. [References] The reference 'V owels et al.' should be 'Vowels et al.'
  3. [Tables 1, 2, and 3] Table 3 appears before Table 2 in the text; renumber the tables in order of first mention.
  4. [Appendix E] The SVI-DAG SVGD kernel is described as an RBF kernel on relaxed DAG-mask features, but no bandwidth or kernel scale is reported; specify this hyperparameter.
  5. [Figure 3] The stacked-bar percentages are not keyed to the legend in a way that is easy to read; label the segments directly or provide the values in a table.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ELBO, priors, and benchmarks are externally grounded; disclosed approximations affect evidence strength, not circularity.

full rationale

Walked the derivation chain. The ELBO (Eq. 3, Appendix A.3) is obtained from the stated joint model p(D|A,θ)p(θ|A)p(B|γ)p(r)p(γ) with a standard change of variables; the Logistic-Beta prior p(γ) (Derivation A.4) is a change of variables from the Beta prior, not the claimed posterior. The SVGD update (Eq. 6) is applied to a target p̃_φ(r|D) = p(r)exp{L(φ;r)}/Z that the paper explicitly identifies as lower-bound-induced and equal to the exact p(r|D) only when Δ_φ(r) is constant; this is a disclosed approximation. Algorithm 1 line 23 uses a straight-through score, and Section 3.6 admits 'The derivative of a sampled hard threshold is not this exact derivative,' so the mode-coverage mechanism is heuristic—a correctness and evidence limitation, not a circular reduction. Section 4.4 warns 'the comparisons presented here should not be taken too seriously,' which weakens the abstract's superiority claim but does not make the benchmarks self-referential; ground-truth DAGs and external baselines are used. The acyclicity representation and Sinkhorn convergence are cited to Annadani et al. (2023), an external group, not to the author's own prior work, so no load-bearing self-citation chain exists. Neither the prior sensitivity study nor the synthetic/real-data evaluations redefine a fitted parameter as a prediction: all metrics compare posterior samples against external ground truth. No step equates a prediction with a fitted input or defines a quantity in terms of the quantity it is supposed to predict.

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

No genuinely new physical entities are introduced. The main hand-tuned inputs are the prior-concentration constants, likelihood noise, KL coefficient, and Sinkhorn temperatures. The methodological novelty is in how existing components are combined.

free parameters (4)
  • Prior concentration constants (nu_min, kappa, eta) = (1, 500, 2)
    Hand-set in Section 3.2.1 and Figure 1 to make the Beta-Bernoulli prior concentrate more strongly away from p=0.5. The concentration controls how much the domain-informed prior resembles a fixed Bernoulli prior and shapes the Logistic-Beta prior over logits.
  • Likelihood noise scale sigma = 0.5 fixed (synthetic ER); learned centered at 0.1 (two-node, Sachs)
    Appendix E: likelihood noise is fixed at 0.5 for all ER benchmarks and learned with a log-noise prior centered at 0.1 elsewhere. The likelihood scale directly changes the expected log-likelihood term in the ELBO.
  • Weight-space KL coefficient = 1.0 (two-node, Sachs); 0.01 (synthetic ER)
    Appendix E: this coefficient scales D_KL(q_phi(theta|A)||p(theta|A)) in the objective and differs across experiment families, so the reported comparisons are not under one fixed objective.
  • Sinkhorn temperature schedule = initial 20 (Sachs), 0.1 elsewhere; final 0.1
    Appendix E: the initial Sinkhorn temperature is experiment-dependent. It controls how close the relaxed order mask is to a hard permutation during training and therefore affects the surrogate gradients.
assumptions (4)
  • domain assumption The data follow an additive noise SCM Xi = f_i(X_Pa(i); theta_i) + epsilon_i with Gaussian noise.
    Section 3.1.1; used for the likelihood p(D|A,theta) and for structural identifiability of the nonlinear ANM.
  • domain assumption The collapsed ELBO in (B,r) space is a valid objective despite being looser than the ELBO in A space.
    Remark 3.1 and Proposition B.2 show D_KL(q(B,r)||p(B,r)) >= D_KL(q(A)||p(A)), so the A-space ELBO is tighter; the paper works in (B,r) for tractability.
  • ad hoc to paper Straight-through surrogate gradients through the hard threshold B=1{gamma+R>0} provide useful optimization signal.
    Assumed for Algorithm 1 and Section 3.4; the paper acknowledges the hard-threshold derivative is not exact and an unbiased estimator would be needed.
  • domain assumption The Sinkhorn relaxation M_tau,KS(r) is sufficiently close to the hard acyclicity mask to make kernel and gradient information meaningful.
    Lemma B.4 gives convergence as tau approaches 0 and KS approaches infinity; finite settings are used in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SVI-DAG: A Structured Variational Inference Approach to Bayesian Causal Discovery." pith.science (2026). https://pith.science/paper/BCGPVSMT

@misc{pith2026260804930,
  author       = {Pith},
  title        = {Pith review of: SVI-DAG: A Structured Variational Inference Approach to Bayesian Causal Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCGPVSMT}},
  note         = {Machine review of arXiv:2608.04930}
}
read the original abstract

Bayesian causal discovery seeks to determine the posterior distribution of causal theories, which are interpreted as directed acyclic graphs (DAGs) that explain the observed data. The resulting posterior allows systematic reasoning regarding epistemic uncertainty within these theories. Nonetheless, finding such graphs is difficult due to identifiability problems and limited observational data. Furthermore, precisely approximating posterior over graphs is challenging given vast range of potential DAGs. Recent Bayesian approaches have addressed some of these challenges, yet they remain limited as they fail to encode dependencies between edges, and lack principled ways to incorporate domain knowledge as inductive biases during the search process. To overcome these limitations, we propose SVI-DAG, a structured variational inference approach to Bayesian causal discovery using observational data and prior beliefs that uses normalizing flows to model dependencies between edges, supporting expressive and multimodal posterior learning over DAGs. To mitigate mode seeking behaviour in evidence lower bound optimization and promote mode coverage, we use stein variational gradient descent to update the node potentials using a kernel in acyclicity space. We evaluate SVI-DAG against 5 state-of-the-art Bayesian DAG learning methods and demonstrate superior performance in uncertainty quantification while remaining competitive in terms of structural accuracy.

Figures

Figures reproduced from arXiv: 2608.04930 by the authors.

Figure 1
Figure 1. Adaptive domain informed prior over adjacency matrix Marginalizing over πij recovers a Beta-Bernoulli model over Bij . Since causal graphs are acyclic, we enforce acyclicity in the prior by construction (Annadani et al., 2023). We introduce a binary free edge matrix B ∈ {0, 1} m×m with zero diagonal and a vector of node poten￾tials r = (r1, . . . , rm) T ∈ R m. Let’s define acyclic mask and induced adjacency by M(r)… view at source ↗
Figure 2
Figure 2. Induced prior over edge logits We now define forward model using Bayesian MLP for node i that uses the parent XPa(i) as input and outputs a prediction Xˆ i . Formally, Xˆ i = fi [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Effect of prior on 2 node graph [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Results on synthetic data generated from linear ER DAGs with [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Results on synthetic data generated from linear ER DAGs with [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Results on synthetic data generated from nonlinear ER DAGs with [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Results on synthetic data generated from nonlinear ER DAGs with [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 5 canonical work pages

  1. [1]

    Ranking via sinkhorn propagation.arXiv preprint arXiv:1106.1925,

    Ryan Prescott Adams and Richard S Zemel. Ranking via sinkhorn propagation.arXiv preprint arXiv:1106.1925,

    arXiv 1925

  2. [5]

    A meta-learning approach to bayesian causal discovery.arXiv preprint arXiv:2412.16577,

    Anish Dhir, Matthew Ashman, James Requeima, and Mark van der Wilk. A meta-learning approach to bayesian causal discovery.arXiv preprint arXiv:2412.16577,

    arXiv

  3. [9]

    The concrete distribution: A continuous relaxation of discrete random variables.arXiv preprint arXiv:1611.00712,

    11 Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables.arXiv preprint arXiv:1611.00712,

    arXiv

  4. [10]

    A graph autoencoder approach to causal structure learning.arXiv preprint arXiv:1911.07420,

    Ignavier Ng, Shengyu Zhu, Zhitang Chen, and Zhuangyan Fang. A graph autoencoder approach to causal structure learning.arXiv preprint arXiv:1911.07420,

    arXiv 1911

  5. [11]

    Masked gradient-based causal structure learning

    Ignavier Ng, Shengyu Zhu, Zhuangyan Fang, Haoyang Li, Zhitang Chen, and Jun Wang. Masked gradient-based causal structure learning. InProceedings of the 2022 SIAM International Conference on Data Mining (SDM), pages 424–432. SIAM,

    work page 2022

  6. [12]

    Bayesian learning of causal structure and mechanisms with gflownets and variational bayes.arXiv preprint arXiv:2211.02763,

    Mizu Nishikawa-Toomey, Tristan Deleu, Jithendaraa Subramanian, Yoshua Bengio, and Laurent Charlin. Bayesian learning of causal structure and mechanisms with gflownets and variational bayes.arXiv preprint arXiv:2211.02763,

    arXiv

  7. [15]

    Advances in variational inference.IEEE transactions on pattern analysis and machine intelligence, 41(8):2008–2026,

    Cheng Zhang, Judith B ¨utepage, Hedvig Kjellstr ¨om, and Stephan Mandt. Advances in variational inference.IEEE transactions on pattern analysis and machine intelligence, 41(8):2008–2026,

    2008

  8. [18]

    Thus the marginal law ofB ν isBernoulli(E[π ν])and converges toBernoulli(p)

    =E[P(B ν = 1|π ν)] =E[π ν] = νp+ 1 ν+ 2 →p. Thus the marginal law ofB ν isBernoulli(E[π ν])and converges toBernoulli(p). Proposition B.2.Let X be a measurable space and let T:X → Ybe a measurable map. Let q and pbe probability measures onXwith pushforwardsq T andp T onY. Then DKL(q∥p)≥D KL(qT ∥p T ),(B.1) with equality if and only if a version of the cond...

    work page 2023

Show all 18 references
  1. [2009]

    Causal discovery with continuous additive noise models.The Journal of Machine Learning Research, 15(1):2009–2053,

    Jonas Peters, Joris M Mooij, Dominik Janzing, and Bernhard Sch ¨olkopf. Causal discovery with continuous additive noise models.The Journal of Machine Learning Research, 15(1):2009–2053,

    2009

  2. [2014]

    Gradient-based neural dag learning.arXiv preprint arXiv:1906.02226,

    S´ebastien Lachapelle, Philippe Brouillard, Tristan Deleu, and Simon Lacoste-Julien. Gradient-based neural dag learning.arXiv preprint arXiv:1906.02226,

    1906 arXiv

  3. [2016]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    arXiv

  4. [2018]

    Differentiable constraint-based causal discovery.arXiv preprint arXiv:2510.22031,

    Jincheng Zhou, Mengbo Wang, Anqi He, Yumeng Zhou, Hessam Olya, Murat Kocaoglu, and Bruno Ribeiro. Differentiable constraint-based causal discovery.arXiv preprint arXiv:2510.22031,

  5. [2019]

    The most computationally intensive components are the formation of the relaxed acyclicity mask and particle interactions in the SVGD style update

    for Tϕ as they provide expressive transformations while allowing tractable forward evaluation and log Jacobian computation. The most computationally intensive components are the formation of the relaxed acyclicity mask and particle interactions in the SVGD style update. For ea...

    2024

  6. [2020]

    Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144,

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144,

    arXiv

  7. [2021]

    Prodag: Projected variational inference for directed acyclic graphs.arXiv preprint arXiv:2405.15167,

    Ryan Thompson, Edwin V Bonilla, and Robert Kohn. Prodag: Projected variational inference for directed acyclic graphs.arXiv preprint arXiv:2405.15167,

  8. [2022]

    Su- perintelligent agents pose catastrophic risks: Can scientist ai offer a safer path?arXiv preprint arXiv:2502.15657,

    Yoshua Bengio, Michael Cohen, Damiano Fornasiere, Joumana Ghosn, Pietro Greiner, Matt Mac- Dermott, S¨oren Mindermann, Adam Oberman, Jesse Richardson, Oliver Richardson, et al. Su- perintelligent agents pose catastrophic risks: Can scientist ai offer a safer path?arXiv preprin...

    arXiv

  9. [2023]

    Variational causal networks: Approximate bayesian inference over causal structures.arXiv preprint arXiv:2106.07635,

    Yashas Annadani, Jonas Rothfuss, Alexandre Lacoste, Nino Scherrer, Anirudh Goyal, Yoshua Bengio, and Stefan Bauer. Variational causal networks: Approximate bayesian inference over causal structures.arXiv preprint arXiv:2106.07635,

    arXiv

  10. [2025]

    International ai safety report 2026.arXiv preprint arXiv:2602.21012,

    Yoshua Bengio, Stephen Clare, Carina Prunkl, Maksym Andriushchenko, Ben Bucknall, Malcolm Murray, Rishi Bommasani, Stephen Casper, Tom Davidson, Raymond Douglas, et al. International ai safety report 2026.arXiv preprint arXiv:2602.21012,

    2026

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.