Research map / Vision-language and multimodal / CLIP and zero-shot
Less-forgetting Multi-lingual Fine-tuning
Yuren Mao, Yaobo Liang, Nan Duan, Haobo Wang, Kai Wang, Lu Chen, Yunjun Gao
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 |
|---|---|---|---|
| fine-tune on multiple languages with uniform weights instead of single-language fine-tuning | XLM-R base · NER zero-shot | F1: 63.97±0.58 -> 67.45±0.35 | Improved |
| fine-tune on multiple languages with uniform weights instead of single-language fine-tuning | XLM-R base · TyDiQA zero-shot | F1: 60.66±0.34 -> 62.51±0.43 | Improved |
| LF-MLF less-forgetting weights instead of uniform weights | multi-lingual fine-tuning · NER zero-shot | F1: 67.45±0.35 -> 68.50±0.21 | Improved |
| LF-MLF instead of uniform weights | multi-lingual fine-tuning · TyDiQA zero-shot | F1: 62.51±0.43 -> 63.03±0.31 | Improved |
| LF-MLF instead of uniform weights | multi-lingual fine-tuning · XNLI zero-shot | F1: 75.17±0.12 -> 75.56±0.15 | Improved |
| LF-MLF instead of PCGrad | multi-lingual fine-tuning · XNLI zero-shot | F1: 74.87±0.21 -> 75.56±0.15 | Improved |
| LF-MLF instead of GradVac | multi-lingual fine-tuning · NER zero-shot | F1: 67.57±0.25 -> 68.50±0.21 | Improved |
| PCGrad or GradVac instead of uniform weights | multi-lingual fine-tuning · NER fine-tuning de | F1: 87.90±0.03 -> 87.13±0.05 | Hurt |
| LF-MLF instead of single-language fine-tuning on source languages | multi-lingual fine-tuning · NER fine-tuning fr | F1: 89.75±0.09 -> 89.15±0.26 | No clear effect |
| LF-MLF instead of uniform weights on source languages | multi-lingual fine-tuning · TyDiQA-GoldP fine-tuning ru | F1: 72.43±0.08 -> 71.66±0.21 | Hurt |
| LF-MLF instead of single-language fine-tuning | multi-lingual fine-tuning · TyDiQA-GoldP fine-tuning en | F1: 66.03±0.31 -> 69.31±0.41 | Improved |
| increase number of source languages from 2 to 10 | LF-MLF and Uniform-MLF · NER zero-shot (Figure 3) | F1 | Improved |
Closest papers by meaning
- Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.87)
- Learn from Downstream and Be Yourself in Multimodal Large Language Models Fine-Tuning (ICML 2025 · similarity 0.87)
- Model Tailor: Mitigating Catastrophic Forgetting in Multi-modal Large Language Models (ICML 2024 · similarity 0.87)
- Enhancing Non-English Capabilities of English-Centric Large Language Models Through Deep Supervision Fine-Tuning (AAAI 2025 · similarity 0.87)
- Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models (NeurIPS 2026 · similarity 0.86)
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 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.
