Research map / Continual learning / Continual fine-tuning of LLMs
Scaling Continual Learning to 300+ Tasks with Bi-Level Routing Mixture-of-Experts
Meng Lou, Yunxiang Fu, Yizhou Yu
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 entropy-driven dynamic per-layer router selection instead of prototype-classifier task identity (ablation) | CaRE · OmniBenchmark-1K 100 tasks | last accuracy AB: 58.29 -> 68.27 | Improved |
| select routers by entropy of class perceptron rather than max logits (ablation) | CaRE · OmniBenchmark-1K 100 tasks | last accuracy AB: 65.21 -> 68.27 | Improved |
| select routers by entropy rather than autoencoder reconstruction error (ablation) | CaRE · OmniBenchmark-1K 100 tasks | last accuracy AB: 64.35 -> 68.27 | Improved |
| select routers by entropy rather than prototype cosine head (ablation) | CaRE · OmniBenchmark-1K 100 tasks | last accuracy AB: 65.91 -> 68.27 | Improved |
| use a separate task-specific router per task instead of one channel-expanded router (ablation) | CaRE · OmniBenchmark-1K 100 tasks | last accuracy AB: 43.96 -> 68.27 | Improved |
| add dynamic router gating over experts instead of unweighted sum of K selected experts (ablation) | CaRE · OmniBenchmark-1K 100 tasks | last accuracy AB: 62.14 -> 68.27 (K=1 w/o gate) | Improved |
| increase K from 1 to 3 without gating (ablation) | CaRE w/o Gate · OmniBenchmark-1K 100 tasks | last accuracy AB: 62.14 -> 61.84 | Hurt |
| use CaRE instead of MOS on 100-task sequence | PTM-based CIL · OmniBenchmark-1K 100 tasks | last accuracy AB: 64.27 -> 68.27 | Improved |
| scale task sequence length to 301 tasks | SEMA and MoAL baselines · OmniBenchmark-1K 301 tasks | last accuracy AB: SEMA 33.01, MoAL 19.33 vs CaRE 68.51 | Hurt |
| go from shallow to deep layers | CaRE routing · OmniBenchmark-1K 301 tasks | Top-2 router recall (%): 27.4 -> 80.6 (layer 3 to 12, T=301) | Improved |
| increase number of learned tasks from 10 to 301 | CaRE routing at layer 12 · OmniBenchmark-1K 301 tasks | Top-2 router recall (%): 92.3 -> 80.6 | Hurt |
Closest papers by meaning
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 2025 · similarity 0.88)
- On Token's Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models (CVPR 2026 · similarity 0.87)
- CoSCL: Cooperation of Small Continual Learners Is Stronger than a Big One (ECCV 2022 · similarity 0.87)
- GACL: Exemplar-Free Generalized Analytic Continual Learning (NeurIPS 2024 · similarity 0.87)
- CLDyB: Towards Dynamic Benchmarking for Continual Learning with Pre-trained Models (ICLR 2025 · similarity 0.87)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.87)
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.
