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..
Soft Task-Aware Routing of Experts for Equivariant Representation Learning
Authors: Jaebyeong Jeon, Hyunseo Jang, Jy-yong Sohn, Kibok Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show consistent improvements across diverse transfer learning tasks. The code is available at https://github.com/YonseiML/star. |
| Researcher Affiliation | Academia | Jaebyeong Jeon Hyeonseo Jang Jy-yong Sohn Kibok Lee Department of Statistics and Data Science, Yonsei University EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual explanations, but it does not include any explicitly labeled pseudocode blocks or algorithm boxes. |
| Open Source Code | Yes | The code is available at https://github.com/YonseiML/star. |
| Open Datasets | Yes | Setup. We pretrain Res Net-18 [19] on STL10 [7] for 200 epochs and Res Net-50 on Image Net100 [43, 45] for 500 epochs, both with a batch size of 256. ... We conduct transfer learning experiments on 11 downstream datasets: CIFAR10/100 [26], Food [2], MIT67 [39], Pets [37], Flowers [34], Caltech101 [13], Cars [25], Aircraft [32], DTD [6], and SUN397 [49]. ... We evaluate our method on the object detection task using the Pascal VOC07+12 dataset [11]. ... We evaluate the generalizability of learned representations under limited data conditions via few-shot classification ... on FC100 [36], CUB200 [48], and Plant Disease datasets [33]. |
| Dataset Splits | Yes | Table B.1: Dataset Information. Overview of the datasets used in the experiments. This table lists dataset names, the number of classes, and the counts for training, validation, and test samples, along with the evaluation metrics. ... CIFAR10 [26] 10 45,000 5,000 10,000 Top-1 accuracy ... Evaluation Protocol: Linear Evaluation. We adopt the standard linear evaluation protocol [3, 17, 24] ... Object Detection. We train a Faster R-CNN [41] with a R50-C4 backbone on the VOC2007+2012 trainval split containing 16551 images. |
| Hardware Specification | Yes | Pretraining on Image Net100 is performed using 4 NVIDIA RTX 4090 GPUs. ... For STL10, pretraining is conducted on a single NVIDIA RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions several frameworks and optimizers like Sim CLR [3], Adam W optimizer [30], cosine learning rate schedule [29], and Faster R-CNN [41], as well as using L-BFGS for optimization. However, it does not provide specific version numbers for these or for core programming languages or libraries (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | Setup. We pretrain Res Net-18 [19] on STL10 [7] for 200 epochs and Res Net-50 on Image Net100 [43, 45] for 500 epochs, both with a batch size of 256. ... The model is trained using the Sim CLR framework [3], employing stochastic gradient descent (SGD) for 500 epochs with a batch size of 256. A cosine annealing learning rate schedule [29] is used, initialized at 0.03 and without restarts, and a weight decay of 0.0005 is applied. The architecture includes 8 experts {Ei}8 i=1, each implemented as a 3-layer MLP with a hidden dimension of 2048 and an output dimension of 128. Batch normalization [22] is excluded from the final layer of each expert. The equivariant predictor ϕT consists of 3 layers, each with a hidden dimension of 512. The routers Rinv and Req are implemented as single-layer MLPs that output 8-dimensional vectors corresponding to the number of experts, followed by softmax activations to produce normalized weights over experts. |