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 in Monotone Mean Field Games

Authors: Noboru Isobe, Kenshi Abe, Kaito Ariu

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments on standard benchmarks confirm that the APP algorithm reliably converges to the unregularized mean-field equilibrium without time-averaging. ... Figure 2 is a summary of the results of the experiment. The most notable aspect is the convergence of exploitability, as shown in Figure 2b. APP decreases the exploitability with each iteration when we update σ. Figure 2a and 2c illustrate the qualitative validity of the approximation achieved by APP.
Researcher Affiliation Collaboration Noboru Isobe RIKEN AIP Tokyo, Japan EMAIL Kenshi Abe Cyber Agent Tokyo, Japan EMAIL Kaito Ariu Cyber Agent Tokyo, Japan EMAIL
Pseudocode Yes Algorithm 1: APP for MFG ... Algorithm 2: RMD(MFG, π0, λ, η, σ0, τ)
Open Source Code No The paper does not contain any explicit statement about releasing source code for the described methodology, nor does it provide a link to a code repository. It mentions that 'The computation of Qλ,σ and µ in Algorithm 1 was based on the implementation provided by Fabian et al. (2023)', referring to the use of a third-party implementation, not their own.
Open Datasets Yes We evaluate the convergence of APP using the Beach Bar Process introduced by Perrin et al. (2020), a standard benchmark for MFGs.
Dataset Splits No The paper describes the parameters of the Beach Bar Process benchmark (e.g., H = 10, |S| = 10, A = { -1, 0, +1}) but does not specify any training, validation, or test dataset splits.
Hardware Specification Yes We ran experiments on a laptop with an 11th Gen Intel Core i7-1165G7 8-core CPU, 16GB RAM, running Windows 11 Pro with WSL.
Software Dependencies No We implemented APP using Python. The computation of Qλ,σ and µ in Algorithm 1 was based on the implementation provided by Fabian et al. (2023). While Python is mentioned, no specific version number for Python or any other libraries/solvers used in their implementation are provided.
Experiment Setup Yes For both algorithms, the learning rate is fixed at η = 0.1, and we vary the regularization parameter λ and update time T to run the experiments. We set H = 10, |S| = 10, A = { -1, 0, +1}, λ = 0.1, η = 0.1, and Ph (s | s, a) = [details of transition kernel] ... In addition, we initialize σ0 and π0 in Algorithm 1 as the uniform distributions on A.