AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Inner-layer Token Self-modulation as Another Scaling Axis for LLMs

Yebin Yang, Huaijin Wu, Jingtao Han, Yu Wang, Xiaohan Qin, Jingzhi Wang, Debing Zhang, Junchi Yan

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: Efficient MoE language model systems

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
add ReToken modulationDense-XL 1.5B · Overall average (14 tasks)accuracy: 22.22 -> 26.54 (+4.32)Improved
add ReToken modulationDense-XL 1.5B · MMLUaccuracy: +4.55Improved
add ReToken modulationDense-XL 1.5B · TriviaQAaccuracy: +9.50Improved
add MoRT1.5B-A250M MoE · Overall average (14 tasks)accuracy: 18.87 -> 22.78 (+3.91)Improved
add MoRT3.2B-A0.5B MoE · Overall average (14 tasks)accuracy: 26.75 -> 32.34 (+5.59)Improved
add MoRT3.2B-A0.5B MoE · ARC-Caccuracy: +7.25Improved
add MoRT3.2B-A0.5B MoE · GSM8Kaccuracy: +6.31Improved
use MoRT (routed pool) instead of ReToken3.2B-A0.5B MoE · Overall average (14 tasks)accuracy: 29.24 -> 32.34Improved
add MoRT at larger backbone scalesMoE 2.3B-A0.5B, 3.9B-A0.6B, A9.8B-A1.4B · Validation lossloss: reduces final loss by 0.059, 0.064, 0.051Improved
add MoRT in isoFLOPs analysisMoE · Compute-optimal lossloss: 2.27% lower loss; 35% compute savingImproved
double the MoRT parameter count (eta sweep 30 to 130) (ablation)MoRT on 3.9B-A0.6B MoE · Validation lossloss: reduces loss by about 0.0118Improved
add embedding parallelism and token deduplication to MoRT (ablation)MoRT on 3.2B-A0.5B MoE · Training throughputtokens/s: 2,749K -> 4,510KImproved

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