Research map / Optimization / Optimizers
Turning Stale Gradients into Stable Gradients: Coherent Coordinate Descent with Implicit Landscape Smoothing for Lightweight Zeroth-Order Optimization
Chen Liang, Xiatao Sun, Qian Wang, Daniel Rakita
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 momentum-based gradient reuse (gamma>0) to block cyclic coordinate descent (ablation) | CoCD · SARCOS | final loss: 188.73 -> 31.18 | Improved |
| add momentum-based gradient reuse (gamma>0) to block cyclic coordinate descent (ablation) | CoCD · MNIST | accuracy (%): 27.03% -> 95.48% | Improved |
| add momentum-based gradient reuse (gamma>0) to block cyclic coordinate descent (ablation) | CoCD · CIFAR-10 | accuracy (%): 10.13% -> 45.08% | Improved |
| add gradient reuse under extreme sparsity (B=1, gamma=0.95) (ablation) | CoCD · SARCOS | final loss: 188.0 -> 68.64 | Improved |
| increase smoothing radius epsilon (ablation) | CoCD · SARCOS | validation loss | Improved |
| increase smoothing radius epsilon to 1 (ablation) | BCCD · SARCOS | validation loss | Hurt |
| increase smoothing radius epsilon (ablation) | CoCD · MNIST | accuracy / stability | Mixed |
| increase momentum gamma (ablation) | CoCD · SARCOS | convergence speed | Improved |
| increase compute budget B beyond 64 (ablation) | CoCD · SARCOS MLP | validation loss: B=64 minimum, larger budgets diminishing returns | No clear effect |
| reduce memory budget M to 0.25 (ablation) | CoCD · SARCOS | validation loss | No clear effect |
| use CoCD instead of ZO-SGD | CoCD · SARCOS (matched 64 evals) | wall-clock per episode (s): 15.7s -> 8.1s | Improved |
| use CoCD instead of SPSA and ZO-SGD on regression | CoCD · SARCOS | stability | Improved |
Closest papers by meaning
- Adaptively Incorporating Directional Hints into Zeroth-Order Optimization (NeurIPS 2026 · similarity 0.89)
- Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations (ICLR 2025 · similarity 0.87)
- Refining Adaptive Zeroth-Order Optimization at Ease (ICML 2025 · similarity 0.87)
- Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization (NeurIPS 2026 · similarity 0.86)
- Double Variance Reduction: A Smoothing Trick for Composite Optimization Problems without First-Order Gradient (ICML 2024 · similarity 0.86)
- Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning (NeurIPS 2026 · 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.
