Research map / Large language models / Retrieval and knowledge
Decoupling Knowledge from Memorization: Retrieval-augmented Prompt Learning
Xiang Chen, Lei Li, Ningyu Zhang, Xiaozhuan Liang, Shumin Deng, Chuanqi Tan, Fei Huang, Luo Si, Huajun Chen
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 |
|---|---|---|---|
| Use retrieval-augmented prompt learning with neural demonstrations, kNN-guided training, and kNN prediction | RoBERTa-large prompt learning · 16-shot classification and information extraction benchmarks | task score (accuracy or F1) | Improved |
| Add kNN-based prediction at test time (ablation) | Ours · SST-2, CR, MNLI, QQP, TACRED (16-shot) | accuracy or F1: 93.2 vs 93.9; 91.2 vs 91.9; 70.4 vs 71.1; 73.0 vs 74.0; 38.2 vs 40.7 | Improved |
| Add kNN-guided training (ablation) | Ours · SST-2, CR, MNLI, QQP, TACRED (16-shot) | accuracy or F1: 92.0 vs 93.9; 90.2 vs 91.9; 68.8 vs 71.1; 71.3 vs 74.0; 36.5 vs 40.7 | Improved |
| Add neural demonstrations (ablation) | Ours · SST-2, CR, MNLI, QQP, TACRED (16-shot) | accuracy or F1: 92.4 vs 93.9; 91.0 vs 91.9; 70.1 vs 71.1; 72.7 vs 74.0; 37.9 vs 40.7 | Improved |
| Refresh the knowledge-store asynchronously (ablation) | Ours · SST-2, CR, MNLI, QQP, TACRED (16-shot) | accuracy or F1: 93.5 vs 93.9; 91.5 vs 91.9; 70.7 vs 71.1; 73.6 vs 74.0; 39.9 vs 40.7 | Improved |
| Use prompt-based keys rather than [CLS] keys (ablation) | Ours · CR and TACRED (16-shot) | accuracy or F1: 89.0 -> 91.9; 37.2 -> 40.7 | Improved |
| Use representation similarity rather than BM25 for kNN acquisition (ablation) | Ours · CR and TACRED (16-shot) | accuracy or F1: 89.5 -> 91.9; 38.8 -> 40.7 | Improved |
| Increase the kNN-training weight β excessively (ablation) | Ours · CR (16-shot) | performance | Hurt |
| Increase kNN neighborhood size (ablation) | Ours · MR (zero-shot) | performance | Improved |
| Increase kNN neighborhood size (ablation) | Ours · MR (16-shot) | performance | No clear effect |
| Interpolate kNN predictions with the prompt model at inference (ablation) | Ours · MR (16-shot) | performance | Mixed |
Closest papers by meaning
- A Similarity Paradigm Through Textual Regularization Without Forgetting (AAAI 2025 · similarity 0.86)
- From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025 · similarity 0.86)
- InstructRetro: Instruction Tuning post Retrieval-Augmented Pretraining (ICML 2024 · similarity 0.86)
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.86)
- Detecting, Explaining, and Mitigating Memorization in Diffusion Models (ICLR 2024 · similarity 0.85)
- Memorization Sinks: Isolating Memorization during LLM Training (ICML 2025 · 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.
