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..

Direct Fisher Score Estimation for Likelihood Maximization

Authors: Sherman Khoo, Yakun Wang, Song Liu, Mark Beaumont

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on a range of synthetic and real-world problems demonstrate the superior performance of our method compared to existing benchmarks.
Researcher Affiliation Academia School of Mathematics, University of Bristol School of Biological Sciences, University of Bristol Correspondence to: Sherman Khoo <EMAIL>
Pseudocode Yes Algorithm 1 FSM-MLE Algorithm (SGD)
Open Source Code Yes Code is available at: https://github.com/Shermjj/Direct_FSM
Open Datasets Yes We train a GAN model on a 16 16 MNIST dataset
Dataset Splits No The paper mentions "N independent and identically distributed observations D = {xi}N i=1" and in Appendix A.11.5, "The gradient comparison experiment corresponding to the plots of Figures 3 and 8 was carried out for a bivariate Gaussian mean model with 10 observations." and "The multivariate Gaussian parameter estimation accuracy in Figure 11 was performed with 100 observations", but does not specify training/test/validation splits or ratios.
Hardware Specification Yes All experiments in this section were performed on a standard consumer laptop, an Intel i7-11370H CPU with 64GB of RAM. [...] An RTX 4090 GPU with 24GB of VRAM, 41GB of RAM was used in this experiment.
Software Dependencies No The paper mentions "implemented in Python and the JAX package" and refers to optimizers like "Adam [Kingma and Ba, 2015]" and "RMSProp [Tieleman, 2012]", and the "SBI package [Boelts et al., 2025]". However, no specific version numbers are provided for Python, JAX, or the SBI package, which is required for a positive answer.
Experiment Setup Yes For the FSM-based estimation, the (σ, η) hyperparameters, corresponding to the proposal variance and step size, were tuned in the exact same way as the KDE-SP gradient hyperparameters (using the prediction error), but over a grid of [10 3, 10 2, 10 1] [10 2, 10 1, 100] instead. The Adam [Kingma and Ba, 2015] optimizer was used for the FSM-based estimation, with averaging over the last 50 iterations of the parameter iterates.