acceptodds
Under review as a conference paper at ICLR 2027

Linesearch-Free Nonlinearly Preconditioned GD with Semiautomatic Geometry Design

Abstract

Nonlinear preconditioning makes it possible to adapt gradient updates to the growth of the objective's curvature, but combining its design with suitable stepsize selection is challenging. We propose *adaptive dual preconditioned gradient descent* (adaptive DPGD), which combines curvature-based preconditioner design with a linesearch-free stepsize rule based on local information. For convex objectives, we establish linear convergence under local strong convexity of the objective. By adapting the stepsize rule, we extend the method to weakly convex objectives and establish asymptotic stationarity without global Lipschitz smoothness. Moreover, our local assumptions give greater freedom in preconditioner design. We exploit this freedom to develop a semiautomatic recipe guided by the objective's curvature, yielding a family of preconditioners that includes those underlying normalized and hyperbolic gradient descent. Experiments on convex and weakly convex problems with real data demonstrate the effectiveness of the preconditioner design and adaptive stepsize selection.

Primary area:
optimization
Keywords:
nonlinear preconditioning, linesearch-free adaptive stepsizes, local relative smoothness, weakly convex optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qpVJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper propose adaptive dual preconditioned gradient descent (DPGD), which combines nonlinear preconditioning with a stepsize rule based on local Bregman curvature estimates. For convex objectives, the authors prove convergence and, under additional local strong convexity assumptions on the preconditioner, R-linear convergence of both objective gaps and gradient norms. They also develop a separate stepsize rule for weakly convex objectives with a known weak convexity parameter, proving asymptotic stationarity without global Lipschitz smoothness. A second contribution is a semiautomatic procedure for constructing radial preconditioners from the normal or tangential inverse curvatures of target growth models. Experiments on eight real datasets compare preconditioner choices and stepsize strategies in terms of gradient evaluations and runtime, supplemented by exponent ablations. The implementation uses a linesearch to initialize the stepsize. Subsequent updates are linesearch-free.

Critical strengths
  1. The convergence result makes adaptive nonlinear preconditioning applicable under local regularity conditions. The convex analysis establishes R-linear convergence without requiring a global dual relative smoothness constant. The weakly convex extension retains the explicit DPGD update and proves asymptotic stationarity through a different stepsize rule and Lyapunov analysis.

  2. The paper gives a concrete procedure for designing preconditioners and evaluates their contribution separately from stepsize adaptation. The normal and tangential constructions connect the geometry of the update to the curvature growth of a target model. They recover several existing methods and produce additional variants, with Proposition 12 establishing the regularity needed for the convergence results. Comparing the same preconditioners under adaptive and line-search strategies helps identify where the gains come from. Reporting runtime as well as gradient evaluations also accounts for the auxiliary function evaluation required by the adaptive rule.

Critical weaknesses
  1. The contribution of the convex algorithm should be distinguished more explicitly from the existing Bregman adaptive framework. The paper acknowledges that its convex stepsize rule is a gradient-space specialization of B-adaPG. I therefore see the linear convergence analysis, the weakly convex extension, and the preconditioner construction as the main contributions. A direct correspondence between the algorithms, assumptions, and convergence statements would make this distinction easier to assess. See Ou et al., Sections 2.1-2.2 (https://arxiv.org/html/2508.01353v2).

  2. The experiments leave uncertainty about how sensitive the gains are to initialization and scaling. Several convex experiments deliberately start in regions with pronounced nonquadratic curvature, and the powered hinge experiments choose an initial point with a maximum residual of 10. The interpolation scale is fixed at 1. In addition, the Malitsky–Mishchenko baseline starts with a stepsize of 10−1010^{-10}, while Adaptive DPGD uses linesearch initialization. The results support the method’s effectiveness in the tested settings; additional sensitivity checks would strengthen the evidence that the design procedure works reliably across different initial points and objective scales.

Questions for authors
  1. Could you provide an explicit correspondence between the convex algorithm and B-adaPG in gradient space, including the curvature estimates and assumptions? Which parts of Theorems 7 and 9 require new arguments beyond this specialization?

  2. Do the main convex comparisons retain their conclusions with zero initialization or different magnitudes of the nonzero initial points? How do the results change when the Malitsky–Mishchenko baseline is given an initial stepsize selected using information comparable to that used by Adaptive DPGD?

  3. How sensitive is performance to the interpolation scale, which is fixed at 1?

2 comments

We thank the reviewer for the constructive comments and suggestions.

(W1, Q1) Relation to B-adaPG and theoretical contributions.

We have clarified the dual relationship between DPGD and Bregman gradient methods in the "Relation to prior work" paragraph of the Introduction and Appendix F.1. Our convex algorithm can be interpreted as B-adaPG in gradient space, with smooth objective ϕ\phi and reference function f∗f^*.

The condition ϕ∈P\phi\in\mathcal P in Assumption 4(ii) requires ∇f(Rd)⊆int⁡dom⁡ϕ\nabla f(\mathbb R^d)\subseteq\operatorname{int}\operatorname{dom}\phi. If we restrict attention to dom⁡ϕ=Rd\operatorname{dom}\phi=\mathbb R^d, the following correspondence holds.

B-adaPG (Ou et al., 2026)Our convex algorithm
Stepsize rule (2.6)Convex case of Eq. (7)
ℓk\ell_k in Eq. (2.3a)L^kd,0\widehat L_k^{\mathrm d,0}
Λk,δ\Lambda_{k,\delta} in Eq. (2.3b)Rk,δ0\mathcal R_{k,\delta}^{0}
αk\alpha_k in Eq. (2.3c)ζk0\zeta_k^0
Assumption 2.1 and condition (4.4)Ensured by Assumption 4 with η=0\eta=0
Theorem 4.4Theorem 8 (previously Theorem 7)

Appendix B.5 gives a proof of Theorem 8 directly in terms of xkx_k, without using convex conjugacy.

In contrast, Theorem 10 (previously Theorem 9) provides a new R-linear convergence result for ff under the additional assumption that ϕ\phi is locally strongly convex (Assumption 9). The bounds on f(xk)−f⋆f(x_k)-f_\star and ∥∇f(xk)∥\|\nabla f(x_k)\| in Eq. (13), proved in Appendix B.6, are not provided by Ou et al. (2026).

Although the dual interpretation above relates DPGD to Bregman gradient methods, DPGD uses ϕ\phi as a potential for minimizing ff, rather than simply minimizing ϕ\phi. This role of ϕ\phi and the freedom to design ϕ\phi as a convex potential also support the contributions highlighted in your review: Section 5 constructs ϕ\phi satisfying Assumption 4(ii) and Assumption 9 (Proposition 14 (previously Proposition 12)). The convexity of ϕ\phi, retained even when ff is not convex, is key to the weakly convex extension in Section 4.2.

We have also added a new variant for the convex case in Eq. (8), changing only the second argument of the corresponding minimum in Eq. (7). In particular, f∈C2(Rd)f\in C^2(\mathbb R^d) and ∇2f≻0\nabla^2f\succ0 are relaxed to f∈C1(Rd)f\in C^1(\mathbb R^d) and convexity. It exploits the known minimum ϕ(0)\phi(0): unlike the objective gap in a general Bregman minimization problem, ϕ(∇f(xk))−ϕ(0)\phi(\nabla f(x_k))-\phi(0) is directly computable. Using this gap directly in the Lyapunov argument yields the new candidate. The weaker assumptions and the counterparts of Theorems 8 and 10 are stated in Assumption 5 and Theorems 25 and 26, with proofs in Appendix B.7. We also added experiments using this rule, showing favorable convergence behavior of the proposed method relative to the baselines (Appendix E.7, Figure 23).

(W2, Q2, Q3) Sensitivity to initialization, objective scaling, and interpolation radius.

We have added Appendix E.6 (Figures 17–22), covering Malitsky–Mishchenko (MM) initialization, initial points, objective scaling, and the interpolation radius.

MM initialization. The initial value λ0=10−10\lambda_0=10^{-10} follows the setup in Malitsky and Mishchenko (2020, Algorithm 1 and Section 2.1). Their adaptive rule is

\sqrt{1+\theta_{k-1}}\lambda_{k-1}, \frac{\lVert x_k-x_{k-1}\rVert}{2\lVert\nabla f(x_k)-\nabla f(x_{k-1})\rVert} \right), \qquad \theta_k=\frac{\lambda_k}{\lambda_{k-1}}, \qquad \theta_0=+\infty.$$ At the first adaptive update ($k=1$), the first candidate is $+\infty$. Thus, the small initial value does not force the subsequent stepsizes to remain comparably small. We interpreted the question about MM's initial stepsize in Q2 as referring to linesearch initialization, in light of the comparison in W2. We therefore tested MM with Armijo initialization: the proposed method still reaches the stopping criterion with fewer gradient evaluations and less runtime than this MM baseline on all four datasets (Appendix E.6.1, Figure 17). **Initial points.** We compared zero initialization and different magnitudes of the initial point. The proposed method remains competitive across these starts, and larger initial norms highlight the benefits of nonlinear preconditioning over Euclidean Adaptive DPGD and MM (Appendix E.6.2, Figures 18–19). **Objective scaling.** We scaled $f$ by $c\in\lbrace0.1,1,10\rbrace$. The proposed method retains its advantage over Euclidean Adaptive DPGD and GD Armijo, and its objective-gap curves remain similar across the tested scales (Appendix E.6.3, Figures 20–21). **Interpolation radius.** Changing the radius has little effect in some experiments and a larger effect in others. The observed differences are consistent with insufficient nonlinear correction when the radius is too large and excessive correction when it is too small. The original radius of 1, used without tuning, performs well (Appendix E.6.4, Figure 22).

Dear Reviewer qpVJ,

Thank you again for your time and valuable comments. Your feedback has been very helpful in improving our manuscript.

We have carefully considered your feedback and revised the manuscript to address your concerns, as detailed in our responses. In particular, we have clarified the relation to B-adaPG and added experiments on initialization, objective scaling, and the interpolation radius.

Would you mind reviewing our responses and revised manuscript and letting us know if you have any remaining questions or concerns? We would be happy to provide further clarification.

Best regards,
The Authors

Reviewer_CuAvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors build on the gradient descent method with nonlinear dual preconditioning (DPGD) introduced by Maddison et al. (2021). The first main contribution is the adaptive stepsize rule for DPGD. The second main contribution is a "semiautomatic" construction of the preconditioner based on "raidal target function". The efficiency of the proposed algorithms is verified using preliminary numerical experiments.

Critical strengths

The main strengths of the paper are the R-linear convergence result under additional assumptions (Theorem 9), convergence analysis for weakly convex functions (Theorem 11), and promising preliminary experimental results.

Critical weaknesses

The main weaknesses of the paper are:

  • The "semiautomatic" design of preconditioners is highly questionable, and its description is very difficult to understand. It actively uses the term "curvature" in relation to the objective function, yet its definition is missing (is it Hessian, or Hessian norm, or something else?). Consequently, the meaning and the purpose of the function t(x)t(x) is unclear. While the authors state the desired "ideal relation" (line 296), there is no precise mathematical or even intuitive explanation of how the choice of function t(x)t(x) is supposed to approach this relation and "correct the curvature growth"? Even in the ℓp\ell_p-regression setting, the explanation is missing. The normal/tangent choice is also voluntary. Overall, in my opinion, this section can be seen as some opaque family of preconditioners with unclear motivation, not a semiautomatic design.

  • The novelty of Theorem 7 is highly limited. The authors acknowledge some similarities with the work of Ou et al. (2026). However the connection is stronger the the paper presents. In particular, the B-adaPG algorithm of Ou et al. (2026) applied to minimizing ϕ(z)\phi(z) with the Bregman kernel f∗(z)f^*(z) is equivelent to the algorothm of this paper. To see this we need to make a transformation zk=∇f(xk)z_k = \nabla f(x_k). Consequently, Theorem 7 is a special instance of Theorem 4.4 by Ou et al. (2026), provided that dom⁡ϕ=Rd\operatorname{dom}{\phi} = \mathbb{R}^d, which is true for all preconditioners constructed in Section 5. The Lyapunov function in (11) is also coincides with the definition in (3.17) of Ou et al. (2026) under the same transformation.

Questions for authors

n/a

2 comments

We thank the reviewer for the constructive comments and suggestions.

(W1) Motivation and procedure of semiautomatic preconditioner design.

We have rewritten Section 5 to make the motivation and construction clearer. The "Design principle" paragraph introduces the ideal relation ∇2ϕ(∇f(x))≈(∇2f(x))−1\nabla^2\phi(\nabla f(x))\approx(\nabla^2f(x))^{-1}, where "curvature" refers to the Hessians in this relation. If ∇2ϕ(∇f(x))=(∇2f(x))−1\nabla^2\phi(\nabla f(x))=(\nabla^2f(x))^{-1} held for all xx, one update would yield a minimizer. For a general ff, however, finding ϕ\phi satisfying this equality is difficult, since such a construction would effectively encode the solution itself. We therefore simplify the design by modeling ff with t(x)=τ(∥x∥)t(x)=\tau(\|x\|), where τ\tau is an elementary function of one variable reflecting, for example, power or exponential growth of the Hessian (Example 13 illustrates this with objectives from our experiments), and seek ∇2ϕ(∇t(x))≈(∇2t(x))−1\nabla^2\phi(\nabla t(x))\approx(\nabla^2t(x))^{-1}.

We use the normal direction by default (see Section 5, "Derivation"). The tangential construction is included primarily for its theoretical value. Together with the freedom in bb, it allows our construction to encompass a wider range of existing preconditioners, as shown in Table 2.

The "Construction" paragraph now clarifies that "semiautomatic" concerns constructing a suitable preconditioner ϕ\phi and verifying its compatibility with ff. For example, in Bregman gradient methods, one often seeks a Bregman divergence that allows subproblems to be solved in closed form and satisfies a global compatibility condition with ff. In DPGD, one similarly needs to construct ϕ\phi and verify global dual relative smoothness, ∇2ϕ(∇f(x))⪯L(∇2f(x))−1\nabla^2\phi(\nabla f(x))\preceq L(\nabla^2f(x))^{-1} for all x∈Rdx\in\mathbb R^d and some L>0L>0 (Maddison et al., 2021, Proposition 3.4). Such constructions have typically relied on heuristic choices and problem-specific analysis. Using local dual relative smoothness (Definition 3), our design replaces this procedure with the following steps: (i) manually choosing tt, (ii) manually choosing b>0b>0, and (iii) determining aa from tt as illustrated in Example 13 and computing ϕ\phi using Eq. (20). Under the conditions on ff in Proposition 14, the construction ensures that Assumptions 4 and 9 hold.

(W2) DPGD structure and additional convergence guarantees.

We clarify two points: DPGD's theoretical value as a nonlinear preconditioning method, and new contributions beyond B-adaPG enabled by the DPGD structure.

For convex ff, DPGD and Bregman gradient methods (BGM) have a dual relationship. For example, Kim et al. (2024, "Mirror Duality in Convex Optimization," Section 2.1) interpret DPGD as the mirror dual of a Bregman gradient method. We use this duality to specialize B-adaPG to the convex DPGD setting, as explained in the revised "Relation to prior work" paragraph in the Introduction and Appendix F.1. This is why our rule can be interpreted as B-adaPG applied to minimizing ϕ\phi with f∗f^* as the Bregman kernel, as described in the review.

However, DPGD is a method with theoretical guarantees for minimizing the original objective ff through nonlinear preconditioning; it uses ϕ\phi as a potential for convergence, rather than merely minimizing ϕ\phi. The freedom to design this potential while retaining guarantees for minimizing ff is central to the nonlinear preconditioning studied by Maddison et al. (2021) and Oikonomidis et al. (2025a,b). Our adaptive formulation serves the same purpose. We exploit this freedom to construct ϕ\phi compatible with our convergence theory (Section 5, Proposition 14).

In addition, ϕ\phi plays an essential role in our weakly convex extension: ϕ\phi remains convex even when ff is only weakly convex. We use this convexity to construct a Lyapunov function and establish ∇f(xk)→0\nabla f(x_k)\to0 (Section 4.2, Theorem 12 (previously Theorem 11)). This use of the DPGD structure goes beyond the dual interpretation available in the convex setting.

We have added a new variant for the convex case that guarantees convergence under weaker assumptions (Eq. (8)). It changes only the second argument of the corresponding minimum in Eq. (7). In particular, f∈C2f\in C^2 and ∇2f≻0\nabla^2f\succ0 are relaxed to f∈C1f\in C^1 and convexity. In a general BGM problem, the objective gap involves an unknown optimal value. In DPGD, ϕ\phi is chosen with its minimum at 00, so Gk=ϕ(∇f(xk))−ϕ(0)G_k=\phi(\nabla f(x_k))-\phi(0) is directly computable. We exploit this DPGD structure to obtain Eq. (8) (see the paragraph immediately after Assumption 5), providing algorithmic and theoretical contributions beyond B-adaPG. The assumptions and counterparts of Theorems 8 and 10 (previously Theorems 7 and 9) are stated in Assumption 5 and Theorems 25 and 26, with proofs in Appendix B.7. We also added experiments using this rule, showing favorable convergence behavior of the proposed method relative to the baselines (Appendix E.7, Figure 23).

Dear Reviewer CuAv,

Thank you again for your time and valuable comments. Your feedback has been very helpful in improving our manuscript.

We have carefully considered your feedback and revised the manuscript to address your concerns, as detailed in our responses. In particular, we have clarified the semiautomatic preconditioner design, the distinction between DPGD and Bregman gradient methods, and our contributions beyond B-adaPG, including convergence guarantees under weaker assumptions.

Would you mind reviewing our responses and revised manuscript and letting us know if you have any remaining questions or concerns? We would be happy to provide further clarification.

Best regards,
The Authors

Reviewer_zZgUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies adaptive dual preconditioned gradient descent (DPGD), a nonlinear preconditioning framework encompassing several existing gradient methods, including normalized and hyperbolic gradient descent. The authors combine linesearch-free adaptive stepsize methods of Malitsky and Mishchenko (2020) and Ou et al. (2026) with the DPGD framework studied by Maddison et al. (2021) and Oikonomidis et al. (2025). Using local primal–dual Bregman curvature information, they establish convergence guarantees for convex and weakly convex objectives without requiring global dual relative smoothness. They also propose a semiautomatic preconditioner design guided by the objective's curvature and present numerical experiments.

Critical strengths
  1. The paper provides a nontrivial extension of linesearch-free adaptive stepsize methods to the nonlinear DPGD framework. In particular, it establishes convergence guarantees based on local primal–dual Bregman curvature information, without requiring global dual relative smoothness.

  2. The paper establishes R-linear convergence under additional assumptions and provides a systematic construction of nonlinear preconditioners. The numerical experiments demonstrate promising performance.

Critical weaknesses
  1. Restrictive assumptions for convex objectives: Although the paper emphasizes relaxing global assumptions, Assumption 4 requires the Hessian of the objective to be positive definite everywhere in the convex setting. Thus, the analysis does not cover general convex objectives, even for the asymptotic convergence result in Theorem 7.

  2. Clarity of semiautomatic geometry design: Section 5 provides a systematic construction of nonlinear preconditioners, but its main contribution beyond existing curvature-based preconditioning methods is not sufficiently clear. In particular, the authors could better explain how the relaxed local assumptions provide additional flexibility in preconditioner design and what is meant by "semiautomatic."

Questions for authors
  1. Is the positive definiteness assumption on the Hessian essential for Theorem 7, or could the result be extended to general convex objectives?

  2. Could the authors provide more intuition for the roles of L^kd,η\widehat L_k^{d,\eta}, Rk,δηR_{k,\delta}^{\eta}, ζkη\zeta_k^\eta, and mkηm_k^\eta, and explain how they motivate the choice of βk+1\beta_{k+1} in Eq. (7)?

  3. Could the authors clarify the main contribution of the semiautomatic geometry design in Section 5, particularly how the target curvature model τ(r)\tau(r) should be selected and what advantages the construction offers over existing preconditioner design methods?

5 comments

We thank the reviewer for the constructive comments and suggestions.

(W1, Q1) Weaker assumptions for convex convergence.

We have added a variant for the convex case (Eq. (8)) giving counterparts of Theorems 8 and 10 (previously Theorems 7 and 9) under weaker assumptions. In particular, f∈C2(Rd)f\in C^2(\mathbb R^d) and ∇2f≻0\nabla^2f\succ0 are relaxed to f∈C1(Rd)f\in C^1(\mathbb R^d) and convexity. It changes only the second argument of the minimum in the convex case of Eq. (7), while retaining the same Lyapunov function and a similar proof strategy. Assumption 5 states the assumptions, and the paragraph immediately after it explains how Eq. (8) avoids the division by zero that can occur in the original rule (7) when ∇2f≻0\nabla^2f\succ0 is relaxed to convexity of ff. The convergence guarantees are summarized at the end of Section 4.1 and proved in Theorems 25 and 26 (Appendix B.7).

In Theorem 26, we formulate the additional assumptions for the R-linear bounds in Eq. (13) as conditions on every compact set for simplicity of presentation. As explained in Remark 27 (Appendix B.7), it suffices that ff be strongly convex with Lipschitz continuous gradient near x⋆x_\star, that the comparison in Eq. (3) hold there, and that ϕ\phi be strongly convex near 00.

The added experiments on the four convex problems again show favorable convergence behavior of the proposed method relative to the baselines (Appendix E.7, Figure 23).

(Q2) Interpretation of the local estimates and stepsize rule.

We have added the requested intuition in Section 3.1, immediately after Eq. (5) for the local estimates and after Eq. (7) for the stepsize rule.

(W2, Q3) Semiautomatic preconditioner design.

We have rewritten Section 5 to clarify how the local conditions in Assumption 4(iii) give additional freedom in choosing ϕ\phi for ff. The convergence guarantees for DPGD in Maddison et al. (2021) are established under global dual relative smoothness of ϕ\phi with respect to ff. In the C2C^2 convex setting with ∇2f≻0\nabla^2f\succ0, this requires ∇2ϕ(∇f(x))⪯L(∇2f(x))−1\nabla^2\phi(\nabla f(x))\preceq L(\nabla^2f(x))^{-1} for all x∈Rdx\in\mathbb R^d and some L>0L>0 (Maddison et al., 2021, Proposition 3.4). Finding ϕ\phi compatible with ff has typically relied on heuristic choices followed by problem-specific a posteriori analysis to verify this inequality for each objective ff. In contrast, for convex ff, our local dual relative smoothness condition in Eq. (3) follows from C2C^2 regularity of ff and ϕ\phi. This allows us to use ∇2ϕ(∇f(x))≈(∇2f(x))−1\nabla^2\phi(\nabla f(x))\approx(\nabla^2f(x))^{-1} as a design guide without imposing the global inequality on each choice of ff and ϕ\phi, as explained in the "Design principle" paragraph.

The revised "Construction" paragraph of Section 5 makes the procedure explicit: (i) manually choose t(x)=τ(∥x∥)t(x)=\tau(\|x\|) to model the growth of ff, (ii) manually choose b>0b>0, and (iii) determine aa from tt as illustrated in Example 13 and compute ϕ\phi using Eq. (20). For example, power or exponential terms in ff suggest corresponding elementary functions τ\tau; Example 13 in Section 5 illustrates this choice using objectives from our numerical experiments and derives the corresponding values of aa. As explained in the "Derivation" paragraph, we use ∇2ϕ(∇t(x))≈(∇2t(x))−1\nabla^2\phi(\nabla t(x))\approx(\nabla^2t(x))^{-1} to reduce the design to one-dimensional calculations. Under the conditions on ff in Proposition 14, the construction ensures that Assumptions 4 and 9 hold. This is the intended meaning of "semiautomatic."

Dear Reviewer zZgU,

Thank you again for your time and valuable comments. Your feedback has been very helpful in improving our manuscript.

We have carefully considered your feedback and revised the manuscript to address your concerns, as detailed in our responses. In particular, we have added a variant for the convex case under weaker assumptions and clarified the semiautomatic preconditioner design.

Would you mind reviewing our responses and revised manuscript and letting us know if you have any remaining questions or concerns? We would be happy to provide further clarification.

Best regards,
The Authors

Thank you for the detailed response and revisions. The new convex result addresses my main theoretical concern, and the revised explanation makes the preconditioner construction clearer, although I still find the term “semiautomatic” somewhat unclear. I will keep my positive assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.