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..
C$^2$Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning
Authors: Kunlun Xu, Yibo Feng, Jiangmeng Li, Yongsheng Qi, Jiahuan Zhou
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on multiple FCL benchmarks demonstrate that C2Prompt achieves state-of-the-art performance. |
| Researcher Affiliation | Academia | Kunlun Xu Wangxuan Institute of Computer Technology Peking University Beijing, China EMAIL Yibo Feng* Wangxuan Institute of Computer Technology Peking University Beijing, China EMAIL Jiangmeng Li University of Chinese Academy of Sciences Beijing, China EMAIL Yongsheng Qi Inner Mongolia University of Technology Hohhot, Inner Mongolia Autonomous Region EMAIL Jiahuan Zhou Wangxuan Institute of Computer Technology Peking University Beijing, China EMAIL |
| Pseudocode | Yes | The overall process of our C2Prompt is shown in Algorithm 1. |
| Open Source Code | Yes | Our source code is available at https://github.com/zhoujiahuan1991/NeurIPS2025-C2Prompt |
| Open Datasets | Yes | We conduct the experiments on three widely used benchmarks in FCL, i.e., Image Net-R[74], Domain Net[75] and CIFAR-100[76]. |
| Dataset Splits | Yes | Image Net-R consists of 30,000 images from 200 categories... The dataset is divided into a training set with 24,000 images and a test set with 6,000 images, and 20% of the training set is selected as a validation set for tuning model parameters. CIFAR-100 contains 50,000 training and 10,000 test-colored images for 100 classes, respectively. For Image Net-R, each task randomly selects 20 classes (20% samples per class), distributed randomly across clients with varying round durations. For Domain Net, each task randomly selects 35 classes (2% samples per class due to closeness to pre-trained distribution, others same as Image Net-R). For each task, training data is distributed among 10 clients following a Dirichlet distribution with parameter β {0.5, 0.1, 0.05} to simulate non-IID scenarios. |
| Hardware Specification | Yes | All experiments are conducted on a single Nvidia 4090 GPU. |
| Software Dependencies | No | The paper mentions "Vi T-B/16 pre-trained on Image Net-21k" as the backbone network and "Adam optimizer" but does not specify version numbers for key software libraries or programming languages. |
| Experiment Setup | Yes | The settings of our discriminativity prompts follow the configuration of previous works [19], where Lp, N and d are set to 10, 8 and 768, respectively. For our class distribution compensation prompt, the prompt length Lc is set to 3 by default. The Adam optimizer with a learning rate of 0.01 is adopted during training. For all the experiments, the training and testing images are resized to 224 224. The client number K and round number for each task are set to 5 and 3, respectively. |