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..
Discrete Neural Flow Samplers with Locally Equivariant Transformer
Authors: Zijing Ou, Ruixiang Zhang, Yingzhen Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we demonstrate the efficacy of DNFS in a wide range of applications, including sampling from unnormalised distributions, training discrete energy-based models, and solving combinatorial optimisation problems. |
| Researcher Affiliation | Collaboration | Zijing Ou1, Ruixiang Zhang2, Yingzhen Li1 1Imperial College London, 2Apple |
| Pseudocode | Yes | The training and sampling procedures are presented in Algorithms 1 and 2. Algorithm 1 Training Procedure of DNFS Algorithm 2 Sampling Procedure of DNFS |
| Open Source Code | Yes | Our code will be released in the open-source git repository. In the interim, we provide the code for sampling from Ising models in the supplementary materials whilst we prepare our code for the full release. |
| Open Datasets | Yes | We further evaluate our method on the task of sampling from the lattice Ising model, which has the form of p(x) exp(x T Jx), x { 1, 1}D, where J = σAD with σ R and AD being the adjacency matrix of a D D grid.2 The adjacency matrix is constructed using A_D = igraph.Graph.Lattice(dim=[D, D], circular=True). Furthermore, for combinatorial optimisation problems: "Specifically, we benchmark MIS on Erd os Rényi (ER) random graphs (Erdos, 1961), comprising 1,000 training and 100 testing instances, each with 16 to 75 vertices." |
| Dataset Splits | Yes | Specifically, we benchmark MIS on Erd os Rényi (ER) random graphs (Erdos, 1961), comprising 1,000 training and 100 testing instances, each with 16 to 75 vertices. |
| Hardware Specification | Yes | All experiments are conducted on a single Nvidia RTX A6000 GPU. |
| Software Dependencies | No | The paper does not explicitly mention any specific software dependencies with version numbers. |
| Experiment Setup | Yes | The model is trained using the Adam W optimizer with a learning rate of 0.0001 and a batch size of 128 for 1,000 epochs (100 steps per epoch). To prevent numerical instability from exploding loss, the log-ratio term log pt(y)/pt(x) is clipped to a maximum value of 5. |