Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Adaptive Riemannian ADMM for Nonsmooth Optimization: Optimal Complexity without Smoothing
Authors: Kangkang Deng, Jiachen Jin, Jiang Hu, Hongxia Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we investigate the numerical performance of the proposed algorithm and report comparative results with existing methods. All experiments are performed in MATLAB R2023b on a 64-bit laptop equipped with Intel i9-13900HX CPU and 32.0 GB RAM. We denote the final objective values as obj and report the CPU time in seconds. All the results are averaged across 10 repeated experiments with random initializations. |
| Researcher Affiliation | Academia | Kangkang Deng College of Science National University of Defense Technology Changsha, CHINA EMAIL, Jiachen Jin College of Science National University of Defense Technology Changsha, CHINA EMAIL, Jiang Hu Yau Mathematical Sciences Center Tsinghua University Beijing, CHINA EMAIL, Hongxia Wang College of Science National University of Defense Technology Changsha, CHINA EMAIL |
| Pseudocode | Yes | Algorithm 1 Adaptive Riemannian ADMM for solving (1). Input: initial point x0, y0, λ0, ρ0, γ0, parameters cρ, cγ. 1: for k = 0, , K 1 do 2: Update auxiliary variable yk+1 via yk+1 = arg min y Rd Lρk(xk, y, λk). (14) 3: Denote Φk(x) := Lρk(x, yk+1, λk) and obtain xk+1 by single gradient step: xk+1 = Rxk( τkgradΦk(xk)). (15) 4: Update the dual step size γk+1 via γk+1 = min γ0 Ax0 y0 log2 2 Axk+1 yk+1 (k + 1)2 log(k + 2), cγ k1/3 log2(k + 1) 5: Update the dual variable λk+1 via λk+1 = λk γk+1(Axk+1 yk+1). (17) 6: end for Output: (x K, y K, λK). |
| Open Source Code | No | Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We will release the code after the review process. |
| Open Datasets | No | The data matrix A Rm n is generated randomly and the entries follow the standard Gaussian distribution. For data generation, the true parameter x is sampled from N(0, Im). and projected onto Sm 1. The features {ai}N i=1 are sampled independently and the labels bi are set to 1 if x ai + ϵi > 0, where noise ϵi N(0, σ2), and -1, otherwise. Here Y Rn p is a given matrix. |
| Dataset Splits | No | All the results are averaged across 10 repeated experiments with random initializations. The problems described (Sparse PCA, Regularized Linear Classifier, Robust Subspace Recovery) are optimization problems on generated data, not typically involving fixed train/test splits. The text does not mention any specific percentages or sample counts for dataset splits. |
| Hardware Specification | Yes | All experiments are performed in MATLAB R2023b on a 64-bit laptop equipped with Intel i9-13900HX CPU and 32.0 GB RAM. |
| Software Dependencies | Yes | All experiments are performed in MATLAB R2023b on a 64-bit laptop equipped with Intel i9-13900HX CPU and 32.0 GB RAM. |
| Experiment Setup | Yes | For ARADMM, we set γ0 = cγ = 50, ρ0 = 5, cρ = 1 and cτ = 0.2. For OADMM, we set ξ = 0.1 and other parameters are the same as their originals. For the other three algorithms, we follow the same settings as [35], where the parameters are optimized through grid searches. All algorithms use identical random initializations and terminate when |F(Xk+1) F(Xk)| 10 8 or after 500 iterations. |