Research map / Continual learning / Continual fine-tuning of LLMs
Memory as a Markov Matrix: Sample Efficient Knowledge Expansion via Token-to-Dictionary Mapping
Kaustubh Pethkar, Ziyang Xiong, Zuofeng Shang, Yingcong Li
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 |
|---|---|---|---|
| embedding tuning (train only new token embedding) instead of full fine-tuning | Embedding tuning (ET) · Synthetic vocabulary, Qwen2.5-7B (N=1000) | WikiText forgetting (loss increase): +7.37 -> .00 | Improved |
| embedding tuning instead of LoRA | Embedding tuning (ET) · Synthetic vocabulary, Llama3.1-8B (N=1000) | test loss: 7.63 -> 3.79 | Improved |
| embedding tuning instead of prompt tuning | Embedding tuning (ET) · Synthetic vocabulary, Llama3.2-1B (N=1000) | test loss: 6.37 -> 3.50 | Improved |
| prompt tuning beats embedding tuning on Phi-3.5 Mini (data generator model) | Embedding tuning (ET) · Synthetic vocabulary, Phi-3.5 Mini (N=1000) | test loss: 1.87 vs 2.75 | Hurt |
| embedding tuning instead of FFT on multiplication token | Embedding tuning (ET) · Llama-3.2-3B-Instruct arithmetic <spec> operator | accuracy of a<spec>b: 81.4 vs 63.5 for a*b | Improved |
| full fine-tuning on new operator causes addition forgetting | Full fine-tuning · Llama-3.2-3B-Instruct addition a+b | accuracy: 100 -> 0 | Hurt |
| embedding tuning keeps addition accuracy | Embedding tuning (ET) · Llama-3.2-3B-Instruct addition a+b | accuracy: 100 -> 100 | Improved |
| ensemble of multiplication phrasings explains <spec> accuracy (ablation) | Base model prompts · Llama-3.2-3B-Instruct base | accuracy: 63.5 -> 73.2 | Improved |
| embedding tuning instead of FFT for Arabic expansion | Embedding tuning (ET) · Qwen2.5-3B English-to-Arabic | target-language loss: 7.95 -> 2.82 | Improved |
| embedding tuning instead of FFT for Arabic expansion | Embedding tuning (ET) · Qwen2.5-3B English-to-Arabic | English forgetting: +11.95 -> 0.00 | Improved |
| embedding tuning instead of LoRA for German expansion | Embedding tuning (ET) · Qwen2.5-3B English-to-German | target-language loss: 3.29 -> 3.27 | Improved |
| embedding tuning outperforms full fine-tuning in low-data regime | Embedding tuning (ET) · Llama-3.2-3B-Instruct multiplication | accuracy | Improved |
Closest papers by meaning
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.89)
- Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory (NeurIPS 2026 · similarity 0.89)
- Spurious Forgetting in Continual Learning of Language Models (ICLR 2025 · similarity 0.89)
- CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models (AAAI 2025 · similarity 0.89)
- Scalable Language Model with Generalized Continual Learning (ICLR 2024 · similarity 0.88)
- M+: Extending MemoryLLM with Scalable Long-Term Memory (ICML 2025 · similarity 0.88)
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.
