ResearchBenchmarksContinual Learning on 20 Newsgroups (test)Follow27Avg Success RateSPHERE16.619.32224.7May 6, 2026Evaluation ResultsMethodMethodLinksAvg Success RateFinal-task Re(K)SPHEREbackbone=4-layer Trans...backbone=4-layer Transformer classifier with Top-K MoE FFN, penalty_ratio=0.12026.052714.55Fine-tunebackbone=4-layer Trans...backbone=4-layer Transformer classifier with Top-K MoE FFN2026.05172.71