Research map / Optimization / Optimizers
Revisiting Zeroth-Order Hessian Approximation: A Single-Step Policy Optimization Lens
Junbin Qiu, Zhaowei Hong, Renzhe Xu, Yao Shu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add query reuse (N = 4 instead of N = 1) to the proposed estimator (ablation) | Proposed estimator · Black-box adversarial attack (MNIST CNN) | iterations to successful attack: 536 -> 403 | Improved |
| add query reuse (N = 6) to the proposed estimator on synthetic functions (ablation) | Proposed estimator · Hessian error, synthetic functions (d=5000) | Frobenius norm Hessian error | Improved |
| use proposed estimator (N = 1) instead of central-difference estimator | Central-difference Hessian estimator · Hessian error, Quadratic and Rosenbrock (d=5000) | Frobenius norm Hessian error: 8x accuracy improvement | Improved |
| use proposed estimator (N = 1) instead of central-difference estimator | Central-difference Hessian estimator · Hessian error, Styblinski-Tang (d=5000) | Frobenius norm Hessian error: 3.4x | Improved |
| use proposed estimator (N = 4) instead of central-difference estimator | Central-difference Hessian estimator · Hessian error, MNIST CNN Fc2.weight | Frobenius norm Hessian error: 9x error reduction | Improved |
| use proposed Hessian optimizer instead of Vanilla ZOO | Vanilla ZOO · Synthetic optimization (Quadratic, Levy, Rosenbrock, Ackley) | convergence speed: 22x average speedup | Improved |
| use proposed estimator (N = 4) instead of HiZOO | HiZOO · Black-box adversarial attack (MNIST CNN) | iterations to successful attack: 1045 -> 403 | Improved |
| use proposed estimator (N = 4) instead of ZOAR (N = 4) | ZOAR · Black-box adversarial attack (MNIST CNN) | iterations to successful attack: 538 -> 403 | Improved |
| use proposed estimator (N = 1, no reuse) instead of ZOAR (N = 4) | ZOAR · Black-box adversarial attack (MNIST CNN) | iterations to successful attack: 538 -> 536 | No clear effect |
| use HiZOO instead of Vanilla ZOO | Vanilla ZOO · Black-box adversarial attack (MNIST CNN) | iterations to successful attack: 1629 -> 1045 | Improved |
Closest papers by meaning
- Refining Adaptive Zeroth-Order Optimization at Ease (ICML 2025 · similarity 0.88)
- Stochastic Zeroth-Order Optimization under Strongly Convexity and Lipschitz Hessian: Minimax Sample Complexity (NeurIPS 2024 · similarity 0.87)
- Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations (ICLR 2025 · similarity 0.87)
- Zeroth-Order Optimization Finds Flat Minima (NeurIPS 2025 · similarity 0.87)
- Adaptively Incorporating Directional Hints into Zeroth-Order Optimization (NeurIPS 2026 · similarity 0.86)
- On the Optimal Construction of Unbiased Gradient Estimators for Zeroth-Order Optimization (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
