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..
EUGens: Efficient, Unified and General Dense Layers
Authors: Sang Min Kim, Byeongchan Kim, Arijit Sehanobish, Somnath Basu Roy Chowdhury, Rahul Kidambi, Dongseok Shim, Kumar Avinava Dubey, Snigdha Chaturvedi, Min-hwan Oh, Krzysztof M Choromanski
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we outline the experimental setup and evaluate the performance of EUGen. Specifically, we design experiments to answer the following research questions: (RQ1) How effective are EUGens in approximating fully-connected feedforward layers (FFLs)? (RQ2) How well do large-scale neural networks (e.g., Transformers & INRs) with EUGens perform? (RQ3) How much speedup do EUGens achieve in large-scale neural networks? |
| Researcher Affiliation | Collaboration | 1Seoul National University 2Independent 3Google Research 4UNC Chapel Hill 5Google Deep Mind 6Columbia University |
| Pseudocode | No | The paper defines the EUGen layer mathematically and provides theoretical proofs, but it does not contain any explicitly labeled pseudocode or algorithm blocks describing implementation steps in a structured format. |
| Open Source Code | Yes | Our code for the EUGen layer implementation and related experiments can be found at https://github.com/ arijitthegame/EUGen/. |
| Open Datasets | Yes | We pre-train this architecture on 36.8B tokens from Open Web Text [35] dataset... (Section 4.2), Image Net [23] and Places365 [125] datasets (Section 4.3), All tests are conducted on the Synthetic Ne RF dataset [61] (Section D.2) |
| Dataset Splits | Yes | We pre-train this architecture on 36.8B tokens from Open Web Text [35] dataset over 50K iterations. (Section 4.2), All tests are conducted on the Synthetic Ne RF dataset [61] (Section D.2). The use of these standard, well-documented datasets implies their standard splits are used for experimentation. |
| Hardware Specification | Yes | This experiment is run on a single NVIDIA RTX 4090. (Section D.1), The inference time is computed using a single NVIDIA RTX 4090 and an AMD Ryzen 7 7700 8-Core Processor. (Section D.2), We performed pre-training on 4 NVIDIA A6000 GPUs on Open Web Text dataset. (Section D.5) |
| Software Dependencies | No | For these experiments, we use the default configuration of the Pytorch implementation of Ne RF [117]. (Section D.2). While PyTorch is mentioned as a key software component, specific version numbers for PyTorch or other libraries are not provided. |
| Experiment Setup | Yes | We pre-train this architecture on 36.8B tokens from Open Web Text [35] dataset over 50K iterations. (Section 4.2), In our experiments, we set the number of random features to 256, with additional results for the different number of random features provided in Section E.2. (Section D.2, Ne RF Experiments), Table 3: Detailed hyperparameters for EUGen-Vi T experiments on Image Net and Places365 datasets. LR stands for learning rate. Num. layers 12, Num. heads 12, Hidden size 768, MLP dim. 3072, Batch size 4096, Base LR 10-3, Optimizer Adam (Table 3, Section D.3) |