Research map / Architectures and efficiency / Pruning and efficient tuning
A-La-Carte Prompt Tuning (APT): Combining Distinct Data via Composable Prompting
Benjamin Bowman, Alessandro Achille, Luca Zancato, Matthew Trager, Pramuditha Perera, Giovanni Paolini, Stefano Soatto
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Compose prompts trained independently on data shards using structured attention | APT · Image classification datasets, including 20-shard settings | accuracy: APT accuracy at 20 shards: 86.3%; 83.9%; 88.7%; 93.3%; 45.4%; 97.6%; 45.8%; full-data accuracy: 86.2%; 86.6%; 91.7%; 93. | Mixed |
| Add prototype-distance weighting to prompt logits (ablation) | APT · Split CIFAR-100 | average accuracy: 83.63 -> 85.21 | Improved |
| Add prototype-distance weighting to prompt logits (ablation) | APT · CORe50 | average accuracy: 90.89 -> 91.14 | Improved |
| Use structured attention rather than naïvely concatenating prompts (ablation) | APT · Prompt composition experiments | accuracy | Improved |
| Use prompt tuning rather than training only classifier heads (ablation) | APT shard composition · In-domain and out-of-domain datasets | accuracy: in-domain, 20 shards: 90.0% vs. 88.5%; out-of-domain, 20 shards: 45.6% vs. 40.5% | Improved |
| Use average ensembling rather than majority vote (ablation) | APT prompt predictions · Shard ensemble datasets | accuracy: Average ensembling minus majority vote: 3.81% (2 shards); average is reported in the range 1.5-3.8%. | Improved |
| Add per-layer memory tokens instead of using a single shallow prompt without memory (ablation) | Prompt tuning · MIT-67 | accuracy: 86.0% -> 86.2% | Improved |
| Use per-layer memory tokens rather than sharing memory tokens across layers (ablation) | Prompt tuning · MIT-67 | accuracy: 86.5% -> 86.2% | Hurt |
| Replace ImageNet-21k pretraining with the ALBEF visual encoder (ablation) | APT · Image classification datasets, 20 shards | average accuracy: 77.3% -> 67.34% | Hurt |
| Use a fixed pretrained backbone and prompts rather than fine-tuning separate models per shard | Shard ensembles · All seven image classification datasets, 20 shards | average accuracy: 41.46% -> 77.3% | Improved |
| Increase the number of data shards from 2 to 20 | APT · Average across seven image classification datasets | accuracy: 84.2% -> 77.3% | Hurt |
Closest papers by meaning
- CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class Incremental Learning (AAAI 2025 · similarity 0.86)
- DePT: Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning (ICLR 2024 · similarity 0.86)
- System Prompt Optimization with Meta-Learning (NeurIPS 2025 · similarity 0.85)
- BayesPrompt: Prompting Large-Scale Pre-Trained Language Models on Few-shot Inference via Debiased Domain Abstraction (ICLR 2024 · similarity 0.85)
- Understanding Prompt Tuning and In-Context Learning via Meta-Learning (NeurIPS 2025 · similarity 0.85)
- Consistent Prompting for Rehearsal-Free Continual Learning (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
