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..
ECO: Evolving Core Knowledge for Efficient Transfer
Authors: Fu Feng, Yucheng Xie, Ruixiao Shi, Jianlu Shen, Jingq Wang, Xin Geng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that ECO achieves efficient initialization and strong generalization across diverse models and tasks, while significantly reducing computational and memory costs compared to conventional methods. 4 Experiments |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, Southeast University, Nanjing, China 2Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China EMAIL |
| Pseudocode | Yes | Algorithm 1 Mutation of the Learngene, Algorithm 2 Genetic Transfer Learning, Algorithm 3 Update of the Learngene Score |
| Open Source Code | No | The paper's NeurIPS checklist for 'Open access to data and code' (Question 5) states 'Yes' but the justification only mentions public datasets: 'The datasets used in the paper are all public datasets, which are described in the Section 4 and Appendix C.' There is no explicit statement or link provided for the release of the code for the methodology described in the paper. |
| Open Datasets | Yes | Datasets. We conduct evolutionary experiments on three datasets of increasing scale. CIFAR-FS [2] and mini Image Net [64] each contain 100 classes... Image Net-1K [10] contains 1,000 classes... We further evaluate the extracted learngenes on four downstream datasets: Oxford Flowers [46], CUB-200-2011 [66], Stanford Cars [15], and Food-101 [4]. |
| Dataset Splits | Yes | CIFAR-FS [2] and mini Image Net [64] each contain 100 classes, split into 64 for training (Dtrain), 16 for validation (Dval), and 20 for novel evaluation. Image Net-1K [10] contains 1,000 classes, divided into 640, 160, and 200 for the same purposes. |
| Hardware Specification | Yes | All experiments are executed on NVIDIA Ge Force RTX 4090 GPUs, with total computational cost comparable to training a typical medium-scale model. |
| Software Dependencies | No | The paper does not explicitly mention any specific software dependencies with version numbers (e.g., Python version, PyTorch version, or other library versions). |
| Experiment Setup | Yes | Training Details. Evolutionary training is conducted independently across networks to support parallelism. For VGG11 and Res Net12, learngenes evolve over 250 generations, each comprising 20 networks trained for 15 epochs. For Mobile Net V3-Large and Res Net50, evolution proceeds for 100 generations, with 6 networks per generation trained for 5 epochs. All experiments are executed on NVIDIA Ge Force RTX 4090 GPUs, with total computational cost comparable to training a typical medium-scale model. Full hyperparameter configurations are detailed in Appendix C.1. |