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..
Last-Iterate Convergence of Smooth Regret Matching$^+$ Variants in Learning Nash Equilibria
Authors: Linjian Meng, Youzhi Zhang, Zhenxing Ge, Tianyu Ding, Shangdong Yang, Zheng Xu, Wenbin Li, Yang Gao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments on (i) randomly generated two-player zero-sum matrix games with sizes [10, 20, 50], (ii) the normal-form representation of two extensive-form games, Kuhn Poker and Goofspiel, (iii) randomly generated three-player zero-sum polymatrix games with sizes [10, 20, 50], (iv) the games presented in Appendix J.1 and J.2, as well as (v) two Leduc Poker variants. ... The duality gap, rdg(x), is used to evaluate the distance to NE, defined as rdg(x) = P i N maxx i âx i , xi x i . ... We compare smooth RM+ variants (SEx RM+, SPRM+, and SOGRM+) with existing RM+ variants (Ex RM+, PRM+, and RM+), as well as traditional last-iterate convergence OMD based algorithms OGDA, EG, and OG4. ... For each game size, we generate 20 instances and report the average duality gaps with variances. |
| Researcher Affiliation | Collaboration | 1 National Key Laboratory for Novel Software Technology, Nanjing University 2 Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, CAS 3 Microsoft Corporation 4 Jiangsu Key Laboratory of Big Data Security and Intelligent Processing, Nanjing University of Posts and Telecommunications |
| Pseudocode | Yes | K Pseudocode of RM+ Variants Mentioned in This Paper Now, we provide the pseudocode of RM+ variants mentioned in this paper. Specifically, the pseudocode of SEx RM+, SPRM+, and SOGRM+ are shown in Algorithms 1, 2, and 3, respectively. |
| Open Source Code | Yes | Our code is available at https://github. com/menglinjian/Neur IPS-2025-SOGRM. |
| Open Datasets | Yes | The normal-form representations of the two extensive-form games are derived from the open-source code provided by Cai et al. [2025] (https://openreview.net/forum?id=LWe VVPu Ix0¬e Id=4vb VJry MNi&referrer=%5BTasks%5D(%2Ftasks)).... For randomly generated two-player zero-sum matrix and three-player zero-sum polymatrix games, each element of the payoff matrix is uniformly sampled from [ 1, 1]. For each game size, we generate 20 instances and report the average duality gaps with variances. |
| Dataset Splits | Yes | For each game size, we generate 20 instances and report the average duality gaps with variances. |
| Hardware Specification | Yes | All experiments are performed on a machine with an i9-13900K CPU and 128 GB of memory. |
| Software Dependencies | No | The paper mentions tools like 'Nashpy [Knight and Campbell, 2018]' and refers to 'open-source code provided by Cai et al. [2025]' for data derivation, but does not specify any software names with version numbers for the implementation of the algorithms or experimental setup (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For initialization, we set Ξi to 1|X i|/|X i| and 0 for smooth and other RM+ variants, respectively. For OGDA, EG, and OG, the initial strategy is the uniform strategy. ... For all tested algorithm, we use simultaneous updates since to our knowledge, the theoretical analysis of last-iterate convergence is based on simultaneous updates. ... Figures 1, 2, 3, 4, 5, and 6 illustrate performance across various η values (e.g., eta=0.01, eta=0.1, eta=1, eta=10). |