On CIFAR-100, GAP-Prompt reported 89.24% average accuracy and 3.03% forgetting, with a reported 1.27% margin over EvoPrompt. The evaluation used rehearsal-free class-incremental learning: classes were disjoint across tasks, past samples were not stored, and task identity was unknown at inference.
The analysis also measured task querying, the step of identifying which task an image belongs to. After the final task, task-querying accuracy was 55.75% on CIFAR-100, 44.69% on ImageNet-R and 44.30% on CUB-200. Classification accuracy was higher in the reported conditional comparison when the task query was correct than when it was wrong.
Three benchmarks, one task structure
The evaluation covered ImageNet-R, CIFAR-100 and CUB-200. Each dataset was partitioned into 10 disjoint-class tasks: 20 classes per task for ImageNet-R, 10 for CIFAR-100 and 20 for CUB-200.
On ImageNet-R, GAP-Prompt reported 78.72% average accuracy and 3.12% forgetting, compared with average accuracies of 77.14% for CPrompt and 76.83% for EvoPrompt. On CUB-200, it reported 87.29% average accuracy and 3.68% forgetting, near the 88.00% joint-training reference; the paper reports margins of 9.18% over OVOR-Deep and 16.86% over RainbowPrompt.
With the iBOT-1K backbone, the paper reports 79.90% on CIFAR-100 and 69.29% on ImageNet-R, with reported margins of 0.79% and 2.73% over CODA-Prompt. With DINO-1K, GAP-Prompt reported 66.46% on ImageNet-R, versus 63.15% for CODA-Prompt, a reported 3.31% margin.
What GAP-Prompt combines
GAP-Prompt combines instance-conditioned gating, which selects prompt information for each input, with dynamic knowledge fusion of current and historical prompts. It also uses shared prompt distillation for early-layer representations.
The authors' hypothesis was that selectively engaging relevant historical prompt knowledge for each input would capture fine-grained differences within a task and help retain older knowledge.
In a CIFAR-100 component ablation, sequential configurations labeled baseline, +ICG, +DKF and +SPD corresponded to average accuracy and forgetting of 84.78% and 4.56%, 87.82% and 4.05%, 88.94% and 3.73%, and 89.24% and 3.03%, respectively.
Task queries and classification performance
Correct-query classification accuracy exceeded wrong-query accuracy in the reported comparisons: 94.35% versus 81.91% on CIFAR-100, 90.65% versus 68.99% on ImageNet-R and 92.75% versus 82.57% on CUB-200.
On ImageNet-R and CUB-200, over-estimated task queries had higher reported accuracy than under-estimated ones: 71.41% versus 66.38% and 84.32% versus 81.20%, respectively.
Measured efficiency and latency
After 10 tasks, thresholding left fewer active prompt gates than candidate gates: an average of 29.61 active gates out of 60.3 candidate gates on CIFAR-100, 19.18 out of 49.9 on ImageNet-R and 7.06 out of 52.7 on CUB-200. On CIFAR-100, measured latency rose from 19.87 milliseconds per image after two tasks to 30.77 milliseconds after 10 tasks.
In a 10-task CIFAR-100 efficiency comparison, GAP-Prompt used 1.36 million extra stored parameters and 1.36 million extra inference parameters, with 67.39 GFLOPs of computation. RainbowPrompt used 6.44 million stored parameters, 1.85 million inference parameters and 67.46 GFLOPs.
The results still have limits
The main results included reported plus-or-minus values in percentage points: 0.28 for accuracy and 0.44 for forgetting on CIFAR-100, 0.35 and 0.54 on ImageNet-R, and 0.45 and 0.27 on CUB-200. The supplied analysis does not define what those values represent, so their statistical meaning is unclear.
The evidence is limited to the named image-classification benchmarks and the rehearsal-free, disjoint-class task setup used in the study. Whether the reported pattern persists with other architectures, datasets, task orders, task counts or random seeds remains open.
The paper is an arXiv version 1 preprint dated 24 August 2026. The supplied document lists author affiliations but reports no funding statement.
Paper data and sources
Original title: GAP-Prompt: Gated Adaptive Prompting for Efficient Continual Learning
Authors: Trung-Anh Dang, Duy-Cuong Bui, Ngoc-Son Vu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text