Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Grad2Task: Improved Few-shot Text Classification Using Gradients for Task Representation
Jixuan Wang, Kuan-Chieh Wang, Frank Rudzicz, Michael Brudno
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| Add gradient-based task-conditioned adapter modulation to ProtoNet-BN | ProtoNet-BN · Meta-testing datasets (4-, 8-, and 16-shot average) | accuracy: 42.76 ± 4.70 -> 44.11 ± 4.63; 44.90 ± 3.52 -> 46.18 ± 3.36; 46.82 ± 2.78 -> 47.68 ± 2.35 | Improved |
| Add task-conditioned adaptation rather than continue training the base model (ablation) | ProtoNet-BN · Meta-testing datasets (4-, 8-, and 16-shot average) | accuracy: 43.22 ± 4.08 -> 44.11 ± 4.63; 45.26 ± 3.29 -> 46.18 ± 3.36; 46.81 ± 2.72 -> 47.68 ± 2.35 | Improved |
| Replace gradients with average input encoding as the task representation (ablation) | Grad2Task · Meta-testing datasets (4-, 8-, and 16-shot average) | accuracy: 44.11 ± 4.63 -> 43.72 ± 5.20; 46.18 ± 3.36 -> 45.96 ± 3.71; 47.68 ± 2.35 -> 47.31 ± 2.44 | Hurt |
| Replace gradients with average input and textual label encodings (ablation) | Grad2Task · Meta-testing datasets (4-, 8-, and 16-shot average) | accuracy: 44.11 ± 4.63 -> 42.79 ± 5.20; 46.18 ± 3.36 -> 45.63 ± 3.24; 47.68 ± 2.35 -> 47.07 ± 2.42 | Hurt |
| Adapt all token representations instead of only [CLS] (ablation) | Grad2Task · Meta-testing datasets (4-, 8-, and 16-shot average) | accuracy: 44.11 ± 4.63 -> 42.79 ± 5.42; 46.18 ± 3.36 -> 44.96 ± 3.42; 47.68 ± 2.35 -> 45.97 ± 2.71 | Hurt |
| Pretrain the task embedding network on same/different task classification (ablation) | Grad2Task · Meta-testing datasets (4-, 8-, and 16-shot average) | accuracy: 44.11 ± 4.63 -> 43.44 ± 5.27; 46.18 ± 3.36 -> 45.80 ± 3.26; 47.68 ± 2.35 -> 47.78 ± 2.62 | Mixed |
| Generate full adapter parameters instead of scaling and shifting parameters (ablation) | Grad2Task · Meta-testing datasets (4-, 8-, and 16-shot average) | accuracy: 44.11 ± 4.63 -> 42.48 ± 4.63; 46.18 ± 3.36 -> 44.85 ± 3.56; 47.68 ± 2.35 -> 47.02 ± 2.45 | Hurt |
Closest papers by meaning
- Tuning Language Models as Training Data Generators for Augmentation-Enhanced Few-Shot Learning (ICML 2023 · similarity 0.87)
- Multitask Prompted Training Enables Zero-Shot Task Generalization (ICLR 2022 · similarity 0.86)
- AdaBoosting Text Prompts for Vision-Language Models (ECCV 2026 · similarity 0.86)
- Conditionally Adaptive Multi-Task Learning: Improving Transfer Learning in NLP Using Fewer Parameters & Less Data (ICLR 2021 · similarity 0.86)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.86)
- Liberating Seen Classes: Boosting Few-Shot and Zero-Shot Text Classification via Anchor Generation and Classification Reframing (AAAI 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.
