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..
pLSTM: parallelizable Linear Source Transition Mark networks
Authors: Korbinian Pöppel, Richard Freinschlag, Thomas Schmied, Wei Lin, Sepp Hochreiter
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first showcase the theoretical advantages of p LSTM on the synthetic arrow-pointing extrapolation task (see Section 5.1). Then we highlight the benefits of p LSTM for two-dimensional input data on Image Net-1K [Deng et al., 2009, Russakovsky et al., 2015], demonstrating scalability to large-scale datasets (see Sections 5.2 and 5.3). Finally, we illustrate how p LSTM scales to more than two input dimensions on established graph benchmarks (see Section 5.4). |
| Researcher Affiliation | Collaboration | 1 ELLIS Unit Linz, LIT AI Lab, Institute for Machine Learning Johannes Kepler University Linz, Austria 2 Zuse School ELIZA 3 NXAI Gmb H EMAIL |
| Pseudocode | No | The methods are described using equations and textual descriptions in sections 3 and 4, but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks were found. |
| Open Source Code | Yes | The complete code is available at https://github.com/ml-jku/plstm_experiments. |
| Open Datasets | Yes | Experiments on Image Net-1k (see Section 5.2) and on molecular graphs (see Section 5.4) show promising results compared to baselines and good scaling behavior to larger model sizes. |
| Dataset Splits | Yes | For validation, we generate 5120 images in the same resolution and at resolution 384 384 to test for extrapolation capabilities. |
| Hardware Specification | Yes | The models take about one hour of training on a single NVIDIA H100-64GB. |
| Software Dependencies | Yes | We use jax version 0.4.32 and CUDA 12.2. |
| Experiment Setup | Yes | Here, we train for 50 epochs with batch size 128, using learning rates [ 1e-4, 3e-4, 1e-3 ] and report the mean validation curves over five seeds at the best learning rate. |