Research map / Continual learning / Continual fine-tuning of LLMs
Any-SSR: How Recursive Least Squares Works in Continual Learning of Large Language Model
Kai Tong, Kang Pan, Xiao Zhang, Erli Meng, Run He, Yawen Cui, Nuoyan Guo, Huiping Zhuang
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 Any-SSR instead of the best exemplar-free baseline | Llama-2-7B-Chat · TRACE Order1 | OP: 55.69 vs 47.63 (SeqFT) | Improved |
| use Any-SSR instead of replay-based SEEKR | Llama-2-7B-Chat · TRACE | OP: 54.99 -> 55.69 (Order1), 54.69 -> 55.69 (Order2) | Improved |
| add analytic router with isolated LoRA subspaces (ablation) | Any-SSR · TRACE | BWT: -0.00 BWT in both orders vs -2.61 for SEEKR | Improved |
| use analytic (AL) router instead of BP-trained router (ablation) | Any-SSR · TRACE routing accuracy | average accuracy drop per phase: 21.7% average accuracy drop per phase for BP router; 100% on previous tasks for analytic | Improved |
| use separate LoRA adapters with analytic router instead of a single LoRA (ablation) | Any-SSR · TRACE | per-task accuracy | Improved |
| move feature layer from Lf=2 to Lf=4 (E=10000) (ablation) | Any-SSR · TRACE Order1 | OP: 54.79 -> 55.69 | Improved |
| move feature layer from Lf=4 to Lf=6 (E=10000) (ablation) | Any-SSR · TRACE Order1 | OP: 55.69 -> 53.21 | Hurt |
| increase expansion size E from 5000 to 10000 (Lf=4) (ablation) | Any-SSR · TRACE Order1 | OP: 54.89 -> 55.69 | Improved |
| increase expansion size E from 10000 to 15000 (Lf=4) (ablation) | Any-SSR · TRACE Order1 | OP: 55.69 -> 55.69 | No clear effect |
| use Any-SSR instead of SeqFT for general ability | Llama-2-7B-Chat · General ability | GA: 43.50 (-4.27) -> 46.51 (-1.26) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LFPT5 — LFPT5: A Unified Framework for Lifelong Few-shot Language Learning Based on Prompt Tuning of T5 (ICLR 2022) · better on 2 of 2 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
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.90)
- Scalable Language Model with Generalized Continual Learning (ICLR 2024 · similarity 0.88)
- Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models (NeurIPS 2026 · similarity 0.87)
- Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.86)
- Continuous Subspace Optimization for Continual Learning (NeurIPS 2025 · similarity 0.86)
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 2025 · 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.
