Ergodic Generative Flows
Authors: Leo Maxime Brunswic, Mateo Clémente, Rui Heng Yang, Adam Sigal, Amir Rasouli, Yinchuan Li
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate IL-EGFs on toy 2D tasks and real-world datasets from NASA on the sphere, using the KL-weak FM loss. Additionally, we conduct toy 2D reinforcement learning experiments with a target reward using the FM loss. [...] We proceed with experiments wherein the state space S is either a flat torus T2, or sphere S2. [...] On S2, we benchmark EGFs on the earth science volcano dataset (NGDC/WDS, 2025). [...] Table 1. Negative log-likelihood scores of the volcano dataset. |
| Researcher Affiliation | Industry | 1Huawei Technologies Canada, Noah s Ark Laboratories 2Huawei. Correspondence to: Leo Maxime Brunswic <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Ergodic flow: IL training |
| Open Source Code | No | The paper refers to a third-party implementation for comparison: 'The Moser Flow is trained using the implementation provided in the authors Git Hub repository (Rozen, 2022)'. There is no explicit statement or link provided for the authors' own code for Ergodic Generative Flows (EGFs). |
| Open Datasets | Yes | On S2, we benchmark EGFs on the earth science volcano dataset (NGDC/WDS, 2025). [...] NGDC/WDS. Global significant volcanic eruptions database. https://www.ncei.noaa.gov/ access/metadata/landing-page/bin/iso? id=gov.noaa.ngdc.mgg.hazards:G10147, 2025. |
| Dataset Splits | No | The paper mentions 'negative log-likelihood on a validation dataset' and 'recalculating the negative log-likelihood on the training dataset' in Appendix D.1, implying the use of training and validation splits. However, it does not provide specific details on how these splits were performed, such as percentages, sample counts, or the methodology used to create them for reproducibility. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models (e.g., NVIDIA A100, RTX 2080 Ti), CPU models (e.g., Intel Xeon, AMD Ryzen), or detailed computer specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions 'We use the Adam W optimizer' for training but does not provide specific version numbers for any software, libraries, or programming languages used in the implementation. |
| Experiment Setup | Yes | The MLPs are tanh hyperbolic activated and initialized using orthogonal initialization (Saxe et al., 2013; Hu et al., 2020). We use the Adam W optimizer (Kingma & Ba, 2015; Loshchilov & Hutter, 2019) for training. [...] An EGF on S = T2 is built with 16 transformations (8 translations and 2 elements of SLd(Z) together with their inverse). Their MLPs have 5 hidden layers of width 32 to parameterize f and π. [...] The Moser Flow is trained using the implementation provided in the authors Git Hub repository (Rozen, 2022) with the only modification being the model size set to 32x3. [...] The two core MLPs of EGF are of size 256x5, compared to the 512x6 used by Rozen et al. (2021). The learning rate is 1e-3 with an exponential decay to 1e-5 at 3000 epochs of 25 steps. [...] Learning rate is kept at 0.001 (Figure 8 caption). |