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..
Scalable Neural Incentive Design with Parameterized Mean-Field Approximation
Authors: Nathan Corecco, Batuhan Yardim, Vinzenz Thoma, Zebang Shen, Niao He
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our methodology on numerical examples of increasing complexity, using AMID to obtain gradient estimates and ADAM [40] as an update rule on parameters θ. All experiment details, including computational resources, can be found in Appendix F. We also provide reference implementations in JAX and Py Torch4. |
| Researcher Affiliation | Academia | Nathan Corecco Department of Computer Science ETH Zurich EMAIL Batuhan Yardim * Department of Computer Science ETH Zurich EMAIL Vinzenz Thoma Department of Computer Science ETH Zurich & ETH AI Center EMAIL Zebang Shen Department of Computer Science ETH Zurich EMAIL Niao He Department of Computer Science ETH Zurich EMAIL |
| Pseudocode | Yes | Algorithm 1 AMID Input: Update rule F, objective G, T, η, τ, θ, ζ0 1: for t 0, . . . , T do Forward pass 2: ζt+1 = (1 ητ)ζt + ηF(θ, ζt) 3: end for 4: s T +1 = θG(θ, ζT +1), a T = ζG(ζT +1) 5: for t T, . . . , 0 do Backward pass 6: at 1 = (1 ητ)at + ηat ζF(θ, ζt) 7: st 1 = st ηat θF(θ, ζt) 8: end for 9: return s0 |
| Open Source Code | Yes | We evaluate our methodology on numerical examples of increasing complexity, using AMID to obtain gradient estimates and ADAM [40] as an update rule on parameters θ. All experiment details, including computational resources, can be found in Appendix F. We also provide reference implementations in JAX and Py Torch4. [...] We provide code with both JAX and Py Torch implementations, along with instructions to execute the experiments. All experiments are seeded to ensure reproducibility. The main experiments are implemented using the JAX codebase. |
| Open Datasets | No | The paper defines and uses simulated environments and parameters for 'congestion pricing in the classical beach bar MFG' and 'sequential auctions' (e.g., settings (A1)-(A3) with specific H, µ0, αmax, etc.). It does not refer to any external, publicly available datasets by name, link, or citation. |
| Dataset Splits | No | The paper describes various experimental settings (A1)-(A3) and states 'We report the training curves in Figure 2, where we evaluate GTval approx throughout training Tval = 500 for robustness.' While it mentions training and evaluation, it does not specify how any data generated within these simulated environments is explicitly partitioned into training, validation, or test sets in a traditional machine learning sense. The settings themselves serve as the environments for evaluation. |
| Hardware Specification | Yes | We run our experiments on a single NVIDIA H100 GPU with an AMD EPYC 16-core CPU. |
| Software Dependencies | No | The experiments were implemented in JAX and Py Torch, the code is provided in the supplementary material. We implement the adjoint method in JAX. For the Py Torch implementation, some code was adapted from MFGLib[31]. |
| Experiment Setup | Yes | All hyperparameters for the baselines as well as AMID are presented in Table 4. For a fair comparison, we perform a grid search on a range of values for the parameters for all baselines and take the best run after 10 repetitions. In our experiments, AMID is robust to hyperparameter choices while zeroth order methods require some tuning. Parameterizing θ in Mmfa. We parameterize pθ h and sold goods αθ h as residual neural networks sharing a base. The base network, f θ base, has din = H + |A| + 1 inputs consisting of one-hot encoded time vector eh, |A|-dimensional vector of bid distribution ν , and remaining goods at round h denoted rh, (given by αmax Ph 1 h =0 αh ). For the input vector xin Rdin, the base residual network is defined as h(1) = Re LU W (1) xin + b(1) , ybase := h(2) = Re LU W (2) h(1) + b(2) + V (2) xin + c(2) Rdhidden. |