Research map / Continual learning / Domain and prompt-based
LFPT5: A Unified Framework for Lifelong Few-shot Language Learning Based on Prompt Tuning of T5
Chengwei Qin, Shafiq Joty
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 generated pseudo-sample replay and KL label consistency to prompt tuning | Prompt tuning on NER domains · CoNLL03 + OntoNotes | F1: 44.34_{\pm 0.46} -> 47.59_{\pm 2.16} | Improved |
| add generated pseudo-sample replay and KL label consistency to prompt tuning | Prompt tuning on classification domains · AGNews + Amazon + DBPedia + Yahoo | accuracy: 28.47_{\pm 9.65} -> 52.71_{\pm 4.19} | Improved |
| add generated pseudo-sample replay and KL label consistency to prompt tuning | Prompt tuning on summarization domains · CNNDM + WikiHow + XSum | A-RG: 15.67_{\pm 0.24} -> 17.05_{\pm 0.92} | Improved |
| replace generated pseudo samples with randomly selected real replay samples | LFPT5 on NER domains · CoNLL03 + OntoNotes | F1: 47.59_{\pm 2.16} -> 48.72_{\pm 0.9} | Improved |
| replace generated pseudo samples with randomly selected real replay samples | LFPT5 on classification domains · AGNews + Amazon + DBPedia + Yahoo | accuracy: 52.71_{\pm 4.19} -> 67.23_{\pm 1.36} | Improved |
| replace generated pseudo samples with randomly selected real replay samples | LFPT5 on summarization domains · CNNDM + WikiHow + XSum | A-RG: 17.05_{\pm 0.92} -> 17.48_{\pm 0.25} | Improved |
| add KL label-consistency loss rather than omit it (ablation) | LFPT5 on summarization · CNNDM + WikiHow + XSum | A-RG: 16.27_{\pm 0.50} -> 17.11_{\pm 0.59} | Improved |
| increase KL loss weight to 0.40 (ablation) | LFPT5 on summarization · CNNDM + WikiHow + XSum | A-RG: 17.11_{\pm 0.59} -> 16.26_{\pm 0.32} | Hurt |
| use 4 pseudo samples instead of 2 (ablation) | LFPT5 on summarization · CNNDM + WikiHow + XSum | A-RG: 16.28_{\pm 0.95} -> 17.05_{\pm 0.92} | Improved |
| increase pseudo samples beyond 4 (ablation) | LFPT5 on summarization · CNNDM + WikiHow + XSum | A-RG: 17.05_{\pm 0.92} -> 16.80_{\pm 0.52}; 16.74_{\pm 0.45}; 16.79_{\pm 0.64} | Hurt |
| initialize new task prompts from prompts learned on a prior task (ablation) | LFPT5 learning task types · Classification and summarization transfer | accuracy: 84.75 -> 86.00; 67.66 -> 65.01; 22.97 -> 22.80 | Mixed |
| use separate LFPT5 task prompts rather than multitask fine-tuning | LFPT5 on three task types · NER, classification, and summarization | F1: 57.81 -> 63.28; 23.36 -> 23.68; 83.25 -> 84.50 | Improved |
Later papers that compared with this method
4 later accepted papers report LFPT5 as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- Sculpting Subspaces: Constrained Full Fine-Tuning in LLMs for Continual Learning (ICLR 2026 · better on 2 of 2 shared results)
- Any-SSR: How Recursive Least Squares Works in Continual Learning of Large Language Model (ICCV 2025 · better on 2 of 2 shared results)
- Scalable Language Model with Generalized Continual Learning (ICLR 2024 · better on 1 of 1 shared result)
- Progressive Prompts: Continual Learning for Language Models (ICLR 2023 · better on 4 of 4 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- The Surprising Effectiveness of Test-Time Training for Few-Shot Learning (ICML 2025 · similarity 0.88)
- Many-Shot In-Context Learning (NeurIPS 2024 · similarity 0.87)
- Stress-Testing Long-Context Language Models with Lifelong ICL and Task Haystack (NeurIPS 2024 · similarity 0.87)
- Optimization Inspired Few-Shot Adaptation for Large Language Models (NeurIPS 2025 · similarity 0.87)
- All You Need is One: Capsule Prompt Tuning with a Single Vector (NeurIPS 2025 · similarity 0.86)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 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.
