Pith. sign in

REVIEW 1 major objections 6 minor 28 references

Small failure regions can swell exponentially before training settles

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

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

T0 review · glm-5.2

2026-07-09 07:31 UTC pith:UPP2H5DK

load-bearing objection Correct math, honest about its limitations, worth a serious referee. the 1 major comments →

arxiv 2607.07538 v1 pith:UPP2H5DK submitted 2026-07-08 cs.LG math.APstat.ML

Avoiding unsafe sets when training with Langevin Dynamics

Adam M. Oberman This is my paper
classification cs.LG math.APstat.ML
keywords mathcaltrainingtrajectoryburn-inequilibriumfailurelossprobability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper studies a basic safety question: if you train a model with noisy gradient descent, what is the probability that the parameters ever enter a designated failure region — a set of parameters producing unsafe behavior — at any point during training, not just at the end? The author models training as overdamped Langevin dynamics on a smooth, strongly convex loss landscape and proves three layered bounds on the in-set probability. At equilibrium, the failure region's mass is exponentially small in the dimension d. Along the trajectory, a shape-free bound shows this probability relaxes to twice the static value after a burn-in of order d, using only the global spectral gap m of the loss. The paper's central observation is that this burn-in is not an artifact: a failure region sitting on the transport path from initialization to equilibrium can transiently swell by a factor exponential in d, even though its equilibrium mass is tiny. To control this swelling, the author introduces a local relaxation rate lambda_{A_H} defined through the spectral measure of the centered indicator of the failure set. For geometrically isolated (flux-isolated) sets, this rate exceeds the global gap m, shrinking the burn-in proportionally; combined with a maximum-principle ceiling on the density ratio, it caps the trajectory probability uniformly in time. The picture: strong convexity sets how fast training relaxes, but the geometry of the unsafe set determines whether the trajectory bulges through it on the way home.

Core claim

The paper's core discovery is that small equilibrium mass of a failure region does not guarantee small transient probability of visiting it, and the failure mode is geometric and dimensional. The author identifies transient swelling — where a set's probability mass balloons far above both its initial and equilibrium values because it lies on the transport path — as the organizing problem, and shows it is amplified exponentially in high dimensions for angular slices of the equilibrium shell. The mechanism is that global convergence measures (KL, chi-squared, Wasserstein) control total divergence but say nothing about a single set's mass at a single time. The remedy is a local relaxation rate,

What carries the argument

The central machinery is the centered indicator trick: pairing the L2(pi) contraction of the density ratio u_t = nu_t/pi with the centered indicator 1_{A_H} - pi(A_H) rather than the raw indicator, which replaces the second moment pi(A_H) by the variance pi(A_H)(1-pi(A_H)) and makes the bound informative for rare sets. The local relaxation rate lambda_{A_H} is defined via the spectral measure of this centered indicator under the Langevin generator, avoiding the Dirichlet-form Rayleigh quotient (which is vacuous for discontinuous indicators). A maximum-principle ceiling M*pi(A_H), requiring bounded initial density ratio M = ||nu_0/pi||_infty, provides uniform-in-time control from above.

Load-bearing premise

The loss landscape J is assumed to satisfy a global two-sided Hessian bound mI <= nabla^2 J <= LI everywhere, meaning strong convexity and smoothness hold across all of R^d. Real neural network losses are non-convex with multiple minima, and the continuous-time isotropic-noise Langevin SDE is an idealization of discrete-time SGD with anisotropic minibatch noise. The maximum-principle ceiling additionally requires a bounded initial density ratio, which fails for point-mass初始化—

What would settle it

The paper itself provides the falsifier: the Ornstein-Uhlenbeck example with a point-mass start at Q=10 and failure region [4,6] shows transient mass swelling from 0 to ~0.75 against an equilibrium mass of ~10^{-5}, a 10^5-fold overshoot. Both protective hypotheses fail simultaneously (M=infty for point masses, and the set is a shell-slice analogue with lambda_{A_H} not bounded below), demonstrating that the bounds are vacuous exactly where swelling is worst.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 6 minor

Summary. The paper studies the probability that a Langevin training trajectory occupies a designated failure region A_H at any time t, under the assumption that the loss J is m-strongly convex and L-smooth. Three bounds are proved: (1) a static bound showing π(A_H) is exponentially small in dimension d (Theorem 4.2); (2) a shape-free dynamic bound ν_t(A_H) ≤ π(A_H)(1 + √(χ²_0/π(A_H)) e^{-mt}) showing relaxation to twice the static bound after a burn-in of order d (Theorem 5.1); (3) a shape-aware bound replacing the global rate m by a local relaxation rate λ_{A_H} ≥ m defined via the spectral measure of the centered indicator, which for flux-isolated sets shortens the burn-in and, combined with a maximum-principle ceiling, caps the trajectory mass uniformly in time (Theorem 6.1). A worked Ornstein-Uhlenbeck example (Section 7) demonstrates that transient swelling by a factor exponential in d is real and that the burn-in window is necessary.

Significance. The paper addresses a well-posed and practically motivated question: bounding the probability of entering an unsafe set during training, not just at convergence. The synthesis of existing tools (Brascamp-Lieb Poincaré, L²(π) contraction, centered indicator trick) into a two-grade bound is clean and the proofs are correct as stated under the hypotheses. The key technical insight — defining λ_{A_H} through the spectral measure of the centered indicator rather than the Dirichlet-form Rayleigh quotient (which is vacuous for indicators) — is legitimate and well-motivated. The OU and shell-slice examples in Section 7 provide falsifiable confirmation that the swelling phenomenon is real and that both grades of bound are sharp. The paper is transparent about its limitations: the strong convexity assumption, the continuous-time idealization, and the per-family verification requirement for flux-isolation.

major comments (1)
  1. Section 6.1, Definition 1 and the conductance discussion: The paper defines λ_{A_H} = inf supp(μ_ϕ) via the spectral theorem and states that for flux-isolated sets one verifies λ_{A_H} > m 'directly by symmetry or barrier comparison.' However, the paper explicitly declines to assert a universal conductance lower bound λ_{A_H} ≥ h²_{A_H}/2 (end of Section 6.1). This means the connection between the geometric diagnostic (conductance h_{A_H}) and the analytical quantity (λ_{A_H}) is established only case-by-case in Section 6.3. The three families (off-shell ball, far-tail ball, shell-slice) are discussed qualitatively but the actual verification of λ_{A_H} > m for families (a) and (b) is not carried out in detail — only the off-shell ball's orthogonality to linear modes is shown (Remark in Section 6.2), giving λ_{A_H} ≥ 2m for that specific case. For the far-tail ball (family (b)), the Eyri
minor comments (6)
  1. Section 4.1: The combined static bound π(A_H) ≤ e^{-A_eff d/2} uses A_eff = max(A, 4mψ(α)/(Ldσ²)), but the crossover condition and the precise relationship between the two regimes could benefit from a more explicit statement of when each term dominates, perhaps as a displayed inequality.
  2. Section 5.2: The explicit χ²_0 formula for a Gaussian start is stated as an envelope but the derivation is omitted. A one-line pointer to where this closed-form comes from (or a brief derivation) would help the reader.
  3. Section 7.1: The overshoot factor C_bad = exp[(d/2)φ(u)] with φ(u) = u - 1 - log u is stated as following from 'a direct Gaussian computation.' Including the key intermediate step would make this important calculation more verifiable.
  4. Figure 3: The three-panel taxonomy is helpful but the labels (a), (b), (c) in the figure caption do not clearly indicate which panel corresponds to which family. Adding explicit labels on the figure panels would improve readability.
  5. Section 2, Related Work: The positioning relative to [ZLC17] is well-explained, but the relationship to the metastability literature [BEGK04, BGK05] could note more explicitly that those results concern exit from wells (leaving a good set) rather than avoidance of a bad set, which is the dual problem studied here.
  6. The reference [BRG+26] is cited as the motivating application (Scientist AI Predictor). The paper would benefit from a brief clarification of whether the constant C_bad in that work's safety argument corresponds to or is bounded by the quantities defined here, to make the connection concrete.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for a careful and accurate reading of the paper. The referee correctly identifies the central technical contribution (the spectral-measure definition of the local relaxation rate), confirms the proofs are correct under the stated hypotheses, and acknowledges the falsifying examples. The one substantive comment concerns the gap between the geometric diagnostic (conductance) and the analytical quantity (lambda_{A_H}), and the incomplete verification for families (a) and (b) in Section 6.3. We agree that the verification for family (b) is not carried out in detail in the current manuscript and will add it. The off-shell ball (family (a)) is already verified via the orthogonality argument, but we will make this more explicit.

read point-by-point responses

  1. Referee: Section 6.1, Definition 1 and the conductance discussion: The paper defines lambda_{A_H} = inf supp(mu_phi) via the spectral theorem and states that for flux-isolated sets one verifies lambda_{A_H} > m 'directly by symmetry or barrier comparison.' However, the paper explicitly declines to assert a universal conductance lower bound lambda_{A_H} >= h^2_{A_H}/2 (end of Section 6.1). This means the connection between the geometric diagnostic (conductance h_{A_H}) and the analytical quantity (lambda_{A_H}) is established only case-by-case in Section 6.3. The three families (off-shell ball, far-tail ball, shell-slice) are discussed qualitatively but the actual verification of lambda_{A_H} > m for families (a) and (b) is not carried out in detail — only the off-shell ball's orthogonality to linear modes is shown (Remark in Section 6.2), giving lambda_{A_H} >= 2m for that specific case. For the远

    Authors: The referee's observation is accurate. We address each family in turn. (a) Off-shell ball: The verification IS present in the manuscript, in the Remark at the end of Section 6.2. For J(Q) = m/2 ||Q||^2, the nonconstant eigenfunctions of -L are Hermite polynomials, with the slowest modes (eigenvalue m) being the linear coordinates Q_i. For a ball centered at the mode, A_H = {||Q|| < r}, symmetry gives <phi_{A_H}, Q_i>_pi = integral over A_H of Q_i d pi = 0 for every i, so phi_{A_H} is orthogonal to the entire eigenvalue-m subspace and lambda_{A_H} >= 2m. This is a complete verification for family (a), not merely a qualitative claim. We will make this more prominent by cross-referencing it explicitly in Section 6.3. (b) Far-tail ball: The referee is correct that the verification is not carried out in detail. The manuscript states qualitatively that 'hitting times grow exponentially (Eyring-Kramers) and lambda_{A_H} inherits the barrier-induced gap,' but does not provide the explicit calculation. We agree this should be substantiated. In the revision we will add a short calculation for the quadratic loss case J(Q) = 1/2 ||Q||^2, where the far-tail ball A_H = {||Q - Q_0|| < r} with ||Q_0|| >> sqrt(d) has an energy barrier of height ~||Q_0||^2/2. The Eyring-Kramers formula gives the principal eigenvalue of the Dirichlet problem on the complement as growing like exp(||Q_0||^2 / (2 sigma^2)), and since the spectral measure of phi_{A_H} is supported on the Dirichlet spectrum of the restricted operator (the barrier decouples the well from the bulk), lambda_{A_H} inherits this rate. We will write this out explicitly. (c) Shell-slice: The manuscript correctly states lambda_{A_H} -> 0 and Theorem 6.1 is vacuous; no verification is needed here. Regarding the conductance connection: revision: partial

  2. Referee: [continuation of the major comment, which was cut off] ...the paper explicitly declines to assert a universal conductance lower bound lambda_{A_H} >= h^2_{A_H}/2. This means the connection between the geometric diagnostic (conductance h_{A_H}) and the analytical quantity (lambda_{A_H}) is established only case-by-case.

    Authors: We agree with the referee that the absence of a universal conductance-to-spectral-gap inequality for lambda_{A_H} is a genuine limitation, and the manuscript is transparent about it. The reason we decline to assert lambda_{A_H} >= h^2_{A_H}/2 is substantive, not an oversight: the standard Cheeger-type bound relates conductance to the spectral gap of the FULL operator via the Dirichlet form and Rayleigh quotient, but lambda_{A_H} is defined through the spectral measure of the centered indicator phi_{A_H}, which is not in the Dirichlet domain (the gradient of an indicator is a surface measure, giving infinite Dirichlet energy). The standard Cheeger argument does not directly transfer to this spectral-measure quantity. Establishing such a bound would require a different proof technique (e.g., a restricted Poincare inequality with boundary conditions, or a capacitary argument), which is beyond the scope of this paper. We will add a sentence in Section 6.1 making this limitation more explicit and flagging the universal conductance bound as an open problem, so that the case-by-case nature of the verification in Section 6.3 is clearly motivated rather than appearing to be an omission. revision: partial

Circularity Check

0 steps flagged

No circularity found; derivation is self-contained against external benchmarks

full rationale

The paper's three main results (Theorems 4.2, 5.1, 6.1) are derived from standard, externally-sourced mathematical tools. The static bound (Theorem 4.2) uses layer-cake/Fubini and L-smoothness volume arguments — no self-citation. The dynamic bound (Theorem 5.1) uses L²(π) contraction from the Brascamp-Lieb Poincaré inequality [BL76] and Bakry-Émery [BE85], both external and standard; the centered-indicator trick is attributed to [LS93, LV07], also external. The local rate λ_{A_H} (Definition 1) is defined spectrally as inf supp(μ_ϕ) — the bottom of the spectral measure of the centered indicator under the Langevin semigroup. Theorem 6.1 then applies this rate via Cauchy-Schwarz to bound set mass. This is a standard spectral-theoretic conversion, not a renaming: λ_{A_H} is a scalar rate defined independently of the quantity ν_t(A_H) being bounded. The paper is transparent that λ_{A_H} lacks a universal conductance lower bound and must be verified per family (Section 6.1, end). The one self-citation [BRG+26] (which includes Oberman as a co-author) provides motivational context for the safety framing but is not load-bearing for any proof. The OU worked example (Section 7) is exactly solvable and serves as an external check on sharpness. No step in the derivation chain reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The paper has no fitted parameters in the circularity sense: m, L, σ, ψ(α) are properties of the loss and noise, not constants tuned to make the bound work. The prefactor K depends on the sublevel-volume modeling assumption (canonical polynomial form), which is a domain assumption on A_H's geometry. The local relaxation rate λ_{A_H} is a new definition but is a derived spectral quantity, not an invented entity pulled from a hat. The main axioms (strong convexity, Langevin idealization, energy gap) are standard domain assumptions for this type of analysis, explicitly acknowledged as idealizations.

free parameters (5)
  • σ (noise level)
    Set by the optimization (batch size, learning rate). Treated as an input parameter, not fitted to data.
  • m (strong convexity constant)
    Lower Hessian bound of the loss J. A property of the loss landscape, not a fitted parameter.
  • L (smoothness constant)
    Upper Hessian bound of the loss J. A property of the loss landscape, not a fitted parameter.
  • ψ(α) (energy gap)
    The energy gap separating A_H from the minimizer. Determined by the alarm function and threshold α, treated as input.
  • K (prefactor in static bound)
    K = C Γ(η+1)(σ²/2)^η e^{-α̂}. Depends on the sublevel volume form Φ(t) = C(t-ψ(α))^η_+, which is a modeling assumption on the geometry of A_H.
axioms (5)
  • domain assumption The training dynamics are well-approximated by the overdamped Langevin SDE dQ_t = -∇J(Q_t)dt + σdW_t with isotropic noise.
    Section 2, 'The Langevin idealization.' The paper states: 'Isotropy of the noise and the continuous-time limit are idealizations; we take them as given and do not address the discretization or anisotropy gaps.'
  • domain assumption The loss J satisfies the global two-sided Hessian envelope mI ⪯ ∇²J(Q) ⪯ LI for all Q ∈ R^d.
    Section 3, 'Convexity and smoothness.' Required for the Brascamp-Lieb Poincaré inequality (spectral gap m) and the static volume bound (L-smoothness).
  • domain assumption The failure region A_H has a monotone link to the loss: J(Q) ≥ ψ(a_H(Q)) for strictly increasing ψ with ψ(0)=0.
    Section 3, 'The failure region.' Ensures A_H ⊆ {J ≥ ψ(α)}, providing the energy gap that makes A_H rare under π.
  • domain assumption The initial law ν_0 has finite chi-squared divergence χ²_0 < ∞ (and for the ceiling, bounded density ratio M < ∞).
    Section 5.2 and Proposition 5.2. Required for the L²(π) contraction and the maximum-principle ceiling respectively. Fails for point-mass starts.
  • domain assumption The sublevel volume V(t) of A_H has canonical polynomial form Φ(t) = C(t-ψ(α))^η_+ for some C>0, η>0.
    Theorem 4.2. Needed to obtain the explicit exponential-in-d rate. The paper does not argue this holds universally; it is a modeling assumption on the geometry of A_H.
invented entities (1)
  • Local relaxation rate λ_{A_H} independent evidence
    purpose: A set-dependent spectral gap replacing the global Poincaré constant m, used to shorten the burn-in for geometrically isolated failure regions.
    Defined via the spectral measure of the centered indicator (Definition 1), not postulated. It is a property of the existing operator -L and the set A_H. The paper provides a falsifiable handle: for the off-shell ball, symmetry gives λ_{A_H} ≥ 2m; for the shell-slice, λ_{A_H} → 0. These are verified by computation in Section 6.3 and 7.

pith-pipeline@v1.1.0-glm · 19039 in / 3989 out tokens · 440285 ms · 2026-07-09T07:31:26.555527+00:00 · methodology

0 comments
read the original abstract

Training a model with noisy gradient descent can be idealized as overdamped Langevin dynamics on the loss landscape, and a natural safety question is to bound the probability $\nu_t(\mathcal{A}_H) = \mathbb{P}(Q_t \in \mathcal{A}_H)$ that the trajectory lies in a designated failure region $\mathcal{A}_H$. We study this for a smooth, strongly convex loss in $d$ dimensions and a failure region separated from the minimizer by an energy gap. Three bounds emerge. At the end of training, the equilibrium mass $\pi(\mathcal{A}_H)$ is exponentially small in $d$, with a complementary energy-barrier rate when the noise is small. Along the trajectory, a shape-free bound $\nu_t(\mathcal{A}_H) \le \pi(\mathcal{A}_H)(1 + \sqrt{\chi_0^2/\pi(\mathcal{A}_H)}\,e^{-mt})$ shows that the in-set probability relaxes to (twice) the static value after a burn-in time of order $d$, using only the global spectral gap $m$ of the loss. A worked Ornstein-Uhlenbeck example shows this burn-in is necessary: an angular slice of the equilibrium shell can transiently swell by a factor exponential in $d$, even though its equilibrium mass is tiny. To rule such swelling out we introduce a local relaxation rate attached to the failure region, defined through the spectral measure of its centered indicator rather than a Dirichlet-form Rayleigh quotient. For geometrically isolated regions this rate exceeds the global one, shrinking the burn-in proportionally, and combined with a maximum-principle ceiling it caps the trajectory probability uniformly in time. The picture is that strong convexity sets how fast training relaxes, but the shape of the unsafe set decides whether the trajectory bulges through it on the way home.

Figures

Figures reproduced from arXiv: 2607.07538 by Adam M. Oberman.

Figure 1
Figure 1. Figure 1: Transient swelling in the 1D Ornstein-Uhlenbeck process with π = N(0, 1). From the point-mass start ν0 = δ10, the law νt passes through N(5, 3/4) at t ∗ = log 2, placing about 75% of its mass in AH = [4, 6], whose equilibrium mass is π(AH) ≈ 3×10−5 . The inset shows νt(AH) peaking near t ∗ and decaying to π(AH). The dynamic bound comes in two grades. The first (Section 5) assumes only the total equilibrium… view at source ↗
Figure 2
Figure 2. Figure 2: The level-set cap. The equilibrium π spreads over the volume cell ℓ d = (πσ2/L) d/2 (dashed circle), while the failure region AH sits past the threshold {J = ψ(α)} at distance ≥ p 2ψ(α)/L from Pn. Both effects suppress π(AH): volume spreading captures a factor e −Ad/2 , the energy gap contributes the Arrhenius factor e −2ψ/σ2 . the bare Arrhenius rate 2ψ(α)/σ2 shaved by the condition number m/L ∈ (0, 1]. F… view at source ↗
Figure 3
Figure 3. Figure 3: High-dimensional taxonomy of failure regions for the quadratic loss J(Q) = 1 2 ∥Q∥ 2 , whose equilibrium π concentrates on the shell ∥Q∥ ∼ √ d. (a) Off-shell ball at the mode: entropic moat, flux-isolated, λAH ≥ 2m. (b) Far-tail ball: energetic moat (Arrhenius barrier), flux-isolated. (c) Shell-slice: mass flows freely along the shell, λAH → 0, not flux-isolated, only the ceiling Mπ(AH) survives. Theorem 6… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · 4 internal anchors

  1. [1]

    , title =

    Pavliotis, Grigorios A. , title =. 2014 , doi =

    work page 2014

  2. [2]

    2014 , isbn =

    Bakry, Dominique and Gentil, Ivan and Ledoux, Michel , title =. 2014 , isbn =

    work page 2014

  3. [3]

    , title =

    Brascamp, Herm Jan and Lieb, Elliott H. , title =. Journal of Functional Analysis , volume =. 1976 , doi =

    work page 1976

  4. [4]

    S\'eminaire de Probabilit\'es XIX 1983/84 , editor =

    Bakry, Dominique and \'Emery, Michel , title =. S\'eminaire de Probabilit\'es XIX 1983/84 , editor =. 1985 , doi =

    work page 1983

  5. [5]

    and Villani, C\'edric , title =

    Markowich, Peter A. and Villani, C\'edric , title =. Matem\'atica Contempor\^anea , volume =

    work page

  6. [6]

    Communications in Partial Differential Equations , volume =

    Arnold, Anton and Markowich, Peter and Toscani, Giuseppe and Unterreiter, Andreas , title =. Communications in Partial Differential Equations , volume =. 2001 , doi =

    work page 2001

  7. [7]

    Random Structures & Algorithms , volume =

    Lov\'asz, L\'aszl\'o and Simonovits, Mikl\'os , title =. Random Structures & Algorithms , volume =. 1993 , doi =

    work page 1993

  8. [8]

    Random Structures & Algorithms , volume =

    Lov\'asz, L\'aszl\'o and Vempala, Santosh , title =. Random Structures & Algorithms , volume =. 2007 , doi =

    work page 2007

  9. [9]

    and Wibisono, Andre , title =

    Vempala, Santosh S. and Wibisono, Andre , title =. Advances in Neural Information Processing Systems 32 (NeurIPS 2019) , pages =. 2019 , eprint =

    work page 2019

  10. [10]

    and Li, Mufan Bill and Shen, Ruoqi and Zhang, Matthew S

    Chewi, Sinho and Erdogdu, Murat A. and Li, Mufan Bill and Shen, Ruoqi and Zhang, Matthew S. , title =. Proceedings of the 35th Conference on Learning Theory (COLT 2022) , series =. 2022 , eprint =

    work page 2022

  11. [11]

    Chewi, Sinho , title =

    work page

  12. [12]

    Proceedings of the 28th International Conference on Machine Learning (ICML 2011) , pages =

    Welling, Max and Teh, Yee Whye , title =. Proceedings of the 28th International Conference on Machine Learning (ICML 2011) , pages =

    work page 2011

  13. [13]

    and Blei, David M

    Mandt, Stephan and Hoffman, Matthew D. and Blei, David M. , title =. Journal of Machine Learning Research , volume =. 2017 , eprint =

    work page 2017

  14. [14]

    Journal of Machine Learning Research , volume =

    Li, Qianxiao and Tai, Cheng and E, Weinan , title =. Journal of Machine Learning Research , volume =. 2019 , eprint =

    work page 2019

  15. [15]

    Annals of Mathematical Sciences and Applications , volume =

    Hu, Wenqing and Li, Chris Junchi and Li, Lei and Liu, Jian-Guo , title =. Annals of Mathematical Sciences and Applications , volume =. 2019 , eprint =

    work page 2019

  16. [16]

    Safety from Honesty in a Disinterested AI Predictor

    Bengio, Yoshua and Richardson, Oliver and Gaven. Safety from Honesty in a Disinterested. 2026 , eprint =. doi:10.48550/arXiv.2606.29657 , note =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.29657 2026

  17. [17]

    and Hosseinzadeh, Rasa and Zhang, Shunshi , title =

    Erdogdu, Murat A. and Hosseinzadeh, Rasa and Zhang, Shunshi , title =. Proceedings of the 25th International Conference on Artificial Intelligence and Statistics (AISTATS 2022) , series =. 2022 , eprint =

    work page 2022

  18. [18]

    Proceedings of the 30th Conference on Learning Theory (COLT 2017) , series =

    Raginsky, Maxim and Rakhlin, Alexander and Telgarsky, Matus , title =. Proceedings of the 30th Conference on Learning Theory (COLT 2017) , series =. 2017 , eprint =

    work page 2017

  19. [19]

    Proceedings of the 30th Conference on Learning Theory (COLT 2017) , series =

    Zhang, Yuchen and Liang, Percy and Charikar, Moses , title =. Proceedings of the 30th Conference on Learning Theory (COLT 2017) , series =. 2017 , eprint =

    work page 2017

  20. [20]

    The Annals of Probability , volume =

    Menz, Georg and Schlichting, Andr\'e , title =. The Annals of Probability , volume =. 2014 , doi =

    work page 2014

  21. [21]

    Journal of the European Mathematical Society , volume =

    Bovier, Anton and Eckhoff, Michael and Gayrard, V\'eronique and Klein, Markus , title =. Journal of the European Mathematical Society , volume =. 2004 , doi =

    work page 2004

  22. [22]

    Journal of the European Mathematical Society , volume =

    Bovier, Anton and Gayrard, V\'eronique and Klein, Markus , title =. Journal of the European Mathematical Society , volume =. 2005 , doi =

    work page 2005

  23. [23]

    , title =

    Lee, Yin Tat and Vempala, Santosh S. , title =. Annals of Mathematics , volume =. 2024 , doi =

    work page 2024

  24. [24]

    2019 , eprint =

    Hubinger, Evan and van Merwijk, Chris and Mikulik, Vladimir and Skalse, Joar and Garrabrant, Scott , title =. 2019 , eprint =

    work page 2019

  25. [25]

    Skalse, Joar and Howe, Nikolaus H. R. and Krasheninnikov, Dmitrii and Krueger, David , title =. Advances in Neural Information Processing Systems 35 (NeurIPS 2022) , year =. 2209.13085 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  26. [26]

    Barrett, David G. T. and Dherin, Benoit , title =. International Conference on Learning Representations (ICLR 2021) , year =. 2009.11162 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  27. [27]

    On the Origin of Implicit Regularization in Stochastic Gradient Descent

    Smith, Samuel L. and Dherin, Benoit and Barrett, David G. T. and De, Soham , title =. International Conference on Learning Representations (ICLR 2021) , year =. 2101.12176 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  28. [28]

    S\'eminaire de Probabilit\'es , volume =

    Cattiaux, Patrick and Guillin, Arnaud , title =. S\'eminaire de Probabilit\'es , volume =. 2014 , note =

    work page 2014