Research map / Large language models / Language, safety and interpretability
P-Adapters: Robustly Extracting Factual Information from Language Models with Diverse Prompts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| adapt natural-language queries with Prefix P-Adapter instead of using them unchanged | BERT Base · ID | P@1: 0.157 -> 0.425 ± 0.01 | Helped |
| adapt natural-language queries with Prefix P-Adapter instead of using them unchanged | BERT Base · OOD Prompts | P@1: 0.157 -> 0.415 ± 0.01 | Helped |
| adapt natural-language queries with Prefix P-Adapter instead of using them unchanged | BERT Base · OOD Objects | P@1: 0.069 -> 0.193 ± 0.00 | Helped |
| adapt natural-language queries with Prefix P-Adapter instead of using them unchanged | BERT Base · OOD Keyboard Errors | P@1: 0.092 -> 0.326 ± 0.01 | Helped |
| adapt natural-language queries with Prefix P-Adapter instead of using them unchanged | BERT Base · ID | consistency: 0.126 -> 0.656 ± 0.02 | Helped |
| adapt natural-language queries with Prefix P-Adapter instead of using them unchanged | BERT Base · OOD Prompts | consistency: 0.133 -> 0.613 ± 0.02 | Helped |
| add a symmetric KL consistency loss (ablation) | Prefix P-Adapter on BERT Base · ID, OOD Prompts, OOD Objects, and OOD Keyboard Errors | consistency: 0.656 ± 0.023 -> 0.680 ± 0.008; 0.613 ± 0.020 -> 0.645 ± 0.007; 0.588 ± 0.026 -> 0.628 ± 0.005; 0.452 ± 0.031 -> 0.507 ± | Helped |
| add a symmetric KL consistency loss (ablation) | Prefix P-Adapter on BERT Base · ID, OOD Prompts, OOD Objects, and OOD Keyboard Errors | P@1: 0.425 ± 0.011 -> 0.415 ± 0.002; 0.415 ± 0.009 -> 0.408 ± 0.001; 0.193 ± 0.004 -> 0.190 ± 0.002; 0.326 ± 0.012 -> 0.326 ± | Hurt |
| keep the subject embedding unmodified in the Rewrite P-Adapter (ablation) | Rewrite P-Adapter on BERT Base · ID | precision: +13% | Helped |
| keep both subject and [MASK] embeddings unmodified in the Rewrite P-Adapter (ablation) | Rewrite P-Adapter on BERT Base · ID | precision: 4% worse | Helped |
| keep only the [MASK] embedding unmodified (ablation) | Rewrite P-Adapter on BERT Base · ID | precision | No clear effect |
| interpolate the rewritten embeddings with original prompt embeddings using α=0.5 (ablation) | Rewrite P-Adapter on BERT Base · ID | precision | Helped |
Closest papers by meaning
- Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.88)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.88)
- Your Mixture-of-Experts LLM Is Secretly an Embedding Model for Free (ICLR 2025 · similarity 0.88)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.87)
- On the Worst Prompt Performance of Large Language Models (NeurIPS 2024 · similarity 0.86)
- Who Knows the Answer? Finding the Best Model and Prompt for Each Query Using Confidence-Based Search (AAAI 2024 · similarity 0.86)
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.