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..
VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion
Authors: Jaekyun Park, Hye Won Chung
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We rigorously benchmark VIPAMIN on 19 vision tasks covering natural, specialized, and structured recognition using two leading self-supervised backbones (Mo Co-v3 [7] and MAE [21]). Additionally, we assess its effectiveness under data-scarce conditions on five few-shot benchmarks. |
| Researcher Affiliation | Academia | Jaekyun Park School of Electrical Engineering KAIST EMAIL Hye Won Chung School of Electrical Engineering KAIST EMAIL |
| Pseudocode | Yes | Detailed algorithm description for VIPAMIN is introduced in Algorithm 1. Algorithm 1 VIPAMIN |
| Open Source Code | Yes | Our code is available at https://github.com/iamjaekyun/vipamin. |
| Open Datasets | Yes | We evaluate our method on two image classification benchmarks: Visual Task Adaptation Benchmark (VTAB-1k) [73] and Fine-Grained Visual Categorization (FGVC). ... The 19 datasets included in VTAB-1k are summarized in Table S2. For the FGVC few-shot experiments, we sample from the training set of five FGVC datasets: CUB-200-2011 [59], NABirds [56], Stanford Dogs [30], Stanford Cars [15], and Oxford Flowers [46]. |
| Dataset Splits | Yes | For few-shot evaluation, we construct k-shot settings by sampling the training set of five FGVC datasets: CUB-200-2011 [59], NABirds [56], Stanford Dogs [30], Stanford Cars [15], and Oxford Flowers [46]. ... The 19 datasets included in VTAB-1k are summarized in Table S2. ... The FGVC few-shot experiments, we sample from the training set and use the validation set for both hyperparameter tuning and extracting embeddings for the initialization procedures of SPT/rand and VIPAMIN. |
| Hardware Specification | Yes | All experiments were conducted on NVIDIA A6000 GPUs with 40GB of memory. |
| Software Dependencies | No | The implementation of VIPAMIN is based on Py Torch [50], with VTAB-1k datasets loaded via Tensor Flow Datasets [1], following established practices in [20, 66, 70]. |
| Experiment Setup | Yes | Hyperparameter Specification Our hyperparameter settings largely follow the configuration used in [66]. ... The full set of hyperparameters is detailed in Table S4. ... Table S4: Hyperparameter tuning range SPT/rand VIPAMIN Batch size 32 Learning rate scheduler Cosine Decay with Linear Warmup Total epochs 100 Prompt length 100 Optimizer Adam W Learning rate {0.05, 0.1, 0.25, 0.5, 1.0, 2.5, 5.0} Weight decay 0.01 k {2, 8, 32, 128} λ {0.0, 0.5, 1.0} |