[ICLR'26] A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models

🚀 News

• January 31, 2026: Evaluation codes for A-TPT are released.
• January 26, 2026: Paper accepted at ICLR 2026! 🎉
• October 29, 2025: Paper is available on arXiv! 📄

For more details, please feel free to check out our:

This repository provides the official PyTorch implementation of our ICLR 2026 paper:

A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models Authors: Shihab Aaqil Ahamed, Udaya S.K.P. Miriya Thanthrige, Ranga Rodrigo, Muhammad Haris Khan

Our major contributions are summarized as follows:

• We introduce a numerical optimization method, called A-TPT, for better calibration of test-time prompt tuning for VLMs. This resolves the suboptimal performance of existing leading calibration techniques for test-time prompt tuning.
• We introduce novel angular diversity that effectively promotes the diversity among textual features, thereby improving the calibration capabilities of VLMs when $N > |D|$ and $N < |D|$. This is accomplished by maximizing the minimum pairwise angular distance between normalized textual features.
• We conduct extensive experiments to validate the generalizability of our approach on different datasets, including medical datasets, across various baselines. The results show that A-TPT surpasses state-of-the-art methods in calibration performance. We also provide thorough analyses, including theoretical aspects. Moreover, our approach provides superior calibration compared to the zero-shot CLIP model, which reveals improved calibration.

Angular Diversity (AD) vs. Expected Calibration Error (ECE)

Hard Prompt	Tuned Prompt

Comparison with C-TPT and O-TPT

Installation

# Clone this repo
git clone https://github.com/MB-Shihab-Aaqil-Ahamed/A-TPT.git
cd A-TPT

# Create a conda enviroment
conda env create -f environment.yml
conda activate atpt

Datasets

We evaluate our method (A-TPT) on fine-grained and natural distribution shift datasets:

• For fine-grained classification, we consider 11 datasets:

  • ImageNet
  • Flower102
  • OxfordPets
  • SUN397
  • DTD
  • Food101
  • StanfordCars
  • Aircraft
  • UCF101
  • EuroSAT
  • Caltech101

• For natural distribution shift, we consider 4 datasets:

  • ImageNet-V2
  • ImageNet-A
  • ImageNet-R
  • ImageNet-Sketch

Prepare the datasets based on the following GitHub repository TPT.

Experiments

In each of the bash script .sh files, change the {data_root} accordingly. And, you can change the CLIP pretrained backbone by modifying the {arch} parameter to either ‘RN50’ or ‘ViT-B/16’. Also, you can change baselines by modifying the {run_type} to either ‘tpt’ or ‘tpt_ts’ or ‘tpt_atpt’.

1. Baseline (CLIP)

bash scripts/test_baseline.sh {dataset}

2. Test-Time Prompt Tuning (TPT)

# for Fine-grained classification
bash scripts/test_tpt_fg.sh {dataset}

# for natural distribution shift
bash scripts/test_tpt_ds.sh {dataset}

# for temperature scaling experiments, change the run_type to tpt_ts in the .sh file.

3. Ours (A-TPT)

# for Fine-grained classification
bash scripts/test_tpt_atpt_fg.sh {dataset}

# for natural distribution shift
bash scripts/test_tpt_atpt_ds.sh {dataset}

The command line argument {dataset} can be specified as follows: for fine-grained classification datasets, ‘I’, ‘DTD’, ‘Flower102’, ‘Food101’, ‘Cars’, ‘SUN397’, ‘Aircraft’, ‘Pets’, ‘Caltech101’, ‘UCF101’, or ‘eurosat’, and for datasets with natural distribution shifts, ‘V2’, ‘A’, ‘R’, or ‘K’.

Results

Comparison on Fine-grained Datasets

CLIP ViT-B/16 (512-d)

Method	Metric	ImageNet	DTD	Flowers102	Food101	SUN397	Aircrafts	OxfordPets	Caltech101	UCF101	EuroSAT	Stanford Cars	Average
Baseline	Acc.	66.70	44.30	67.30	83.60	62.50	23.90	88.00	92.90	65.00	41.30	65.30	63.70
	ECE	2.12	8.50	3.00	2.39	2.53	5.11	4.37	5.50	3.59	13.89	4.25	4.43
TPT	Acc.	69.00	46.70	69.00	84.70	64.50	23.40	87.10	93.80	67.30	42.40	66.30	65.00
	ECE	10.60	21.20	13.50	3.98	11.30	16.80	5.77	4.51	2.54	13.20	5.16	11.60
C-TPT	Acc.	68.50	46.00	69.80	83.70	64.80	24.85	88.20	93.63	65.70	43.20	65.80	64.57
	ECE	3.15	11.90	5.04	3.43	5.04	4.36	1.90	4.24	2.54	13.20	1.59	5.13
O-TPT	Acc.	67.33	45.68	70.07	84.13	64.23	23.64	87.95	93.95	64.16	42.84	64.53	64.41
	ECE	1.96	7.88	3.87	1.46	4.93	3.68	1.90	3.80	2.34	12.98	1.78	4.23
A-TPT	Acc.	67.70	45.51	69.22	83.64	66.04	23.76	88.33	93.87	66.16	44.06	65.78	64.92
(Ours)	ECE	1.45	4.76	3.61	1.37	3.28	3.14	1.17	2.76	2.12	3.92	1.09	2.61

CLIP RN50 (1024-d)

Method	Metric	ImageNet	DTD	Flowers102	Food101	SUN397	Aircrafts	OxfordPets	Caltech101	UCF101	EuroSAT	Stanford Cars	Average
Baseline	Acc.	58.10	40.00	61.00	74.00	58.60	15.60	83.80	85.80	58.40	23.70	55.70	55.90
	ECE	2.09	9.91	3.19	3.11	3.54	6.45	5.91	4.33	3.05	15.40	4.70	5.61
TPT	Acc.	60.70	41.50	62.50	74.90	61.10	17.00	84.50	87.00	59.50	28.30	58.00	57.70
	ECE	11.40	25.70	13.40	5.25	9.24	16.10	3.65	5.04	12.40	22.50	3.76	11.70
C-TPT	Acc.	60.20	42.20	65.20	74.70	61.00	17.00	84.10	86.90	59.70	27.80	56.50	57.75
	ECE	3.01	19.80	4.14	1.86	2.93	10.70	2.77	2.07	3.83	15.10	1.94	6.19
O-TPT	Acc.	58.97	41.90	65.61	74.22	60.85	16.77	83.40	86.86	58.84	28.35	56.44	57.47
	ECE	3.10	16.53	2.50	1.20	3.20	8.18	3.50	2.75	2.60	14.71	1.69	5.45
A-TPT	Acc.	58.44	40.90	64.89	74.10	60.46	14.58	83.48	86.57	60.24	32.14	57.08	57.53
(Ours)	ECE	2.49	6.41	2.39	1.11	2.90	6.14	2.47	1.98	2.34	2.51	1.38	2.92

Comparison on Natural Distribution Shift Datasets

CLIP ViT-B/16 (512-d)

Method	Metric	ImageNet-A	ImageNet-V2	ImageNet-R	ImageNet-S	Average
Baseline	Acc.	47.80	60.80	74.00	46.10	57.20
	ECE	8.61	3.01	3.58	4.95	5.04
TPT	Acc.	52.60	63.00	76.70	47.50	59.90
	ECE	16.40	11.10	4.36	16.10	12.00
C-TPT	Acc.	51.60	62.70	76.00	47.90	59.60
	ECE	8.16	6.23	1.54	7.35	5.82
O-TPT	Acc.	49.87	61.65	72.55	47.12	57.80
	ECE	7.22	3.97	1.46	6.87	4.88
A-TPT	Acc.	50.39	60.90	74.87	46.09	58.06
(Ours)	ECE	6.45	2.96	1.39	4.87	3.92

CLIP RN50 (1024-d)

Method	Metric	ImageNet-A	ImageNet-V2	ImageNet-R	ImageNet-S	Average
Baseline	Acc.	21.70	51.40	56.00	33.30	40.60
	ECE	21.30	3.33	2.07	3.15	7.46
TPT	Acc.	25.20	54.60	58.90	35.10	43.50
	ECE	31.00	13.10	9.18	13.70	16.70
C-TPT	Acc.	23.40	54.70	58.00	35.10	42.80
	ECE	25.40	8.58	4.57	9.70	12.10
O-TPT	Acc.	23.07	53.11	54.47	33.98	41.16
	ECE	24.56	3.87	4.47	5.85	9.69
A-TPT	Acc.	21.66	51.48	55.78	33.37	40.57
(Ours)	ECE	21.14	3.10	3.96	3.09	7.82

Qualitative Results - Reliability Diagrams

CLIP ViT-B/16

Method	Food	DTD	FLW	Air	UCF	Cars	SUN
C-TPT
O-TPT
A-TPT

CLIP RN50 Backbone

Method	Air	UCF	Cars	SUN
C-TPT
O-TPT
A-TPT

Acknowledgement

The computational resources for this research were supported by the Accelerating Higher Education Expansion and Development (AHEAD) Operation Grant No. 6026-LK/8743-LK from the Ministry of Higher Education, Sri Lanka, funded by the World Bank and the National Research Council of Sri Lanka Grant No. 19-080.

Also we would like thank the authors of the CoOp/CoCoOp, TPT and C-TPT for releasing their code open-source and their instructions for data preparation.

Citation

If you find our work, this repository useful in your research, please consider giving a star ⭐ and citation.

@inproceedings{ahamed2026atpt,
   title     = {A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models},
   author    = {Shihab Aaqil Ahamed and Udaya Sampath K. Perera Miriya Thanthrige and Ranga Rodrigo and Muhammad Haris Khan},
   booktitle = {The Fourteenth International Conference on Learning Representations},
   year      = {2026},
   url       = {https://openreview.net/forum?id=VhlSBZebEw}
}

Contact

If you have any questions, please feel free to reach out at shihabaaqilahamed@gmail.com