---
license: mit
pipeline_tag: image-classification
---

# TALON: Test-time Adaptive Learning for On-the-Fly Category Discovery

TALON is a test-time adaptation framework for on-the-fly category discovery (OCD) that enables a model to recognize known categories while simultaneously discovering novel ones from an unlabeled online stream. 

This repository contains the official implementation and weights for the paper [TALON: Test-time Adaptive Learning for On-the-Fly Category Discovery](https://huggingface.co/papers/2603.08075), presented at **CVPR 2026**.

- **GitHub:** [https://github.com/ynanwu/TALON](https://github.com/ynanwu/TALON)
- **Paper:** [arXiv:2603.08075](https://huggingface.co/papers/2603.08075)

## Method Overview

Existing OCD methods often freeze the feature extractor, which limits the learning potential of incoming data. TALON addresses this with two complementary strategies:
1. **Semantic-aware prototype update**: Dynamically refines class prototypes to improve classification.
2. **Stable test-time encoder update**: Integrates new information directly into the parameter space.
3. **Margin-aware logit calibration**: Applied during the offline stage to reserve embedding space for future class discovery.

## Installation

This project uses [`uv`](https://github.com/astral-sh/uv) for dependency management.

```bash
# Clone the repository
git clone https://github.com/ynanwu/TALON
cd TALON

# Install all dependencies
uv sync
```

## Usage

To evaluate a pretrained checkpoint (e.g., CUB with a CLIP backbone):

```bash
uv run test.py --dataset_name cub --backbone clip --ckpt_path checkpoints/clip/cub/best_model.pth
```

## Citation

If you find this work useful for your research, please consider citing the paper:

```bibtex
@inproceedings{talon2026,
  title={TALON: Test-time Adaptive Learning for On-the-Fly Category Discovery},
  author={Wu, Yanan and Yan, Yuhan and Chen, Tailai and Chi, Zhixiang and Wu, ZiZhang and Jin, Yi and Wang, Yang and Li Zhenbo},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}
```