ReBaPL: Repulsive Bayesian Prompt Learning

CVPR 2026

ReBaPL: Repulsive Bayesian Prompt Learning
Yassir Bendou*, Omar Ezzahir*, Eduardo Fernandes Montesuma*, Gabriel Mahuas, Victoria Shevchenko, Mike Gartrell
* Equal contribution

[Paper]

---

Highlights

ReBaPL is a modular, plug-and-play Bayesian extension for prompt learning methods in vision-language models. It introduces Repulsive Cyclical SGHMC (rcSGHMC), which:

• Uses Hamiltonian dynamics with cyclical learning rates to alternate between exploration and exploitation phases when sampling from complex multimodal posterior distributions of prompts.
• Introduces a representation-based repulsive force derived from probability metrics (MMD and Wasserstein distance) computed on prompt-induced feature distributions, encouraging diverse mode discovery and preventing premature convergence.
• Produces a Bayesian ensemble of prompt samples that provides richer characterization of the posterior, leading to improved generalization on base-to-novel tasks, cross-dataset transfer, and domain generalization.

ReBaPL can extend any existing MLE-based prompt learning method. We demonstrate it on top of MaPLe and MMRL.

---

Installation

This codebase is tested on Ubuntu 20.04 with Python 3.8. Follow the steps below to set up the environment.

1. Create conda environment

conda create -y -n rebapl python=3.8
conda activate rebapl

# Install PyTorch (>= 1.8.1). Adjust the CUDA version to match your system.
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 \
    -f https://download.pytorch.org/whl/torch_stable.html

2. Install the Dassl library

git clone https://github.com/KaiyangZhou/Dassl.pytorch.git
cd Dassl.pytorch/
pip install -r requirements.txt
python setup.py develop
cd ..

3. Clone ReBaPL and install requirements

git clone https://github.com/SigmaNova/ReBaPL.git
cd ReBaPL/
pip install -r requirements.txt
pip install setuptools==59.5.0

4. Data preparation

Please refer to DATASETS.md for instructions on how to install the 11 benchmark datasets. We recommend placing all datasets under a single directory (e.g., ~/datasets).

---

Training and Evaluation

Base-to-Novel Generalization

Quick start using the convenience script:

# Usage: bash run_ours.sh <dataset> <seed> [cfg] [gpu_id] [overwrite]
bash run_ours.sh eurosat 1 vit_b16 0 false

This runs both training on base classes and evaluation on novel classes.

Step-by-step:

# Train on base classes
bash scripts/csghmc_cold_restarts_maple/base2new_train.sh eurosat 1 vit_b16

# Evaluate on novel classes
bash scripts/csghmc_cold_restarts_maple/base2new_test.sh eurosat 1 vit_b16

With custom hyperparameters (e.g., for sweeps):

python train.py \
    --root ~/datasets \
    --seed 1 \
    --trainer CSGHMC_CR_MAPLE \
    --dataset-config-file configs/datasets/eurosat.yaml \
    --config-file configs/trainers/CSGHMC_CR_MAPLE/vit_b16.yaml \
    --output-dir output/base2new/train_base/eurosat/shots_16/CSGHMC_CR_MAPLE/vit_b16/seed1 \
    DATASET.NUM_SHOTS 16 \
    DATASET.SUBSAMPLE_CLASSES base \
    CSGHMC.REPULSION.REPULSION_STRENGTH 0.75 \
    CSGHMC.REPULSION.DISTANCE_TYPE wasserstein \
    OPTIM.MAX_EPOCH 15

Config overrides can be passed as trailing key-value pairs on the command line.

Supported Datasets

ImageNet, Caltech101, OxfordPets, StanfordCars, Flowers102, Food101, FGVCAircraft, SUN397, DTD, EuroSAT, UCF101

Plus OOD variants: ImageNetV2, ImageNet-Sketch, ImageNet-A, ImageNet-R

Parsing Results

python parse_test_res.py output/base2new/test_new/eurosat/shots_16/CSGHMC_CR_MAPLE/vit_b16 --test-log

---

Key Configuration

The main config files are in configs/trainers/CSGHMC_CR_MAPLE/. Below are the hyperparameters used in our experiments (Table 1 of the paper):

Hyperparameter	MaPLe + ReBaPL	MMRL + ReBaPL
Learning rate α	0.002	0.001
Batch size b	1	4
Epochs per cycle T	5	5
# Cycles C	3	3
Samples per cycle K	1	1
Repulsion strength ξ	0.001	10⁻⁴
Repulsion batch size n	32	64
Distance	MMD	Wasserstein
Algorithm	rcSGHMC	rcSGHMC
Weight decay	5e-4	5e-4
Random seed	[1, 2, 3]	[1, 2, 3]

---

Citation

If you find this work useful, please cite our paper:

@article{bendou2025rebapl,
  title={ReBaPL: Repulsive Bayesian Prompt Learning},
  author={Bendou, Yassir and Ezzahir, Omar and Montesuma, Eduardo Fernandes and Mahuas, Gabriel and Shevchenko, Victoria and Gartrell, Mike},
  journal={arXiv preprint arXiv:2511.17339},
  year={2025}
}

Acknowledgements

Our codebase builds on MaPLe, PromptSRC, and Dassl. We thank the authors for releasing their code.