AI Paper Atlas

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

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Low-rank and model-merging adaptation · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use entropy-driven dynamic per-layer router selection instead of prototype-classifier task identity (ablation)CaRE · OmniBenchmark-1K 100 taskslast accuracy AB: 58.29 -> 68.27Improved
select routers by entropy of class perceptron rather than max logits (ablation)CaRE · OmniBenchmark-1K 100 taskslast accuracy AB: 65.21 -> 68.27Improved
select routers by entropy rather than autoencoder reconstruction error (ablation)CaRE · OmniBenchmark-1K 100 taskslast accuracy AB: 64.35 -> 68.27Improved
select routers by entropy rather than prototype cosine head (ablation)CaRE · OmniBenchmark-1K 100 taskslast accuracy AB: 65.91 -> 68.27Improved
use a separate task-specific router per task instead of one channel-expanded router (ablation)CaRE · OmniBenchmark-1K 100 taskslast accuracy AB: 43.96 -> 68.27Improved
add dynamic router gating over experts instead of unweighted sum of K selected experts (ablation)CaRE · OmniBenchmark-1K 100 taskslast 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 taskslast accuracy AB: 62.14 -> 61.84Hurt
use CaRE instead of MOS on 100-task sequencePTM-based CIL · OmniBenchmark-1K 100 taskslast accuracy AB: 64.27 -> 68.27Improved
scale task sequence length to 301 tasksSEMA and MoAL baselines · OmniBenchmark-1K 301 taskslast accuracy AB: SEMA 33.01, MoAL 19.33 vs CaRE 68.51Hurt
go from shallow to deep layersCaRE routing · OmniBenchmark-1K 301 tasksTop-2 router recall (%): 27.4 -> 80.6 (layer 3 to 12, T=301)Improved
increase number of learned tasks from 10 to 301CaRE routing at layer 12 · OmniBenchmark-1K 301 tasksTop-2 router recall (%): 92.3 -> 80.6Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy