AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

UltraMemV2: Memory Networks Scaling to 120B Parameters with Superior Long-Context Learning

Zihao Huang, Yu Bao, Qiyang Min, Siyan Chen, Ran Guo, Hongzhi Huang, Defa Zhu, Banggu Wu, Yutao Zeng, zhou Xun, Siyuan Qiao

ICLR 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization

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 PEER-style FFN values instead of value embeddings (ablation)UltraMemV2 · UltraMemV2-430M/5Btraining loss and Open-Benchmark accuracyImproved
single value projector and separate queries per Tucker rank (ablation)UltraMemV2 · UltraMemV2-430M/5B after 1.5T tokensOpen-Benchmark accuracy: +0.7 points; loss -0.0026Improved
use one head instead of two (ablation)UltraMemV2 · UltraMemV2-430M/5B after 1T tokensOpen-Benchmark accuracy: +0.2 points; loss -8e-4No clear effect
remove Tucker core penalty loss (ablation)UltraMemV2 · UltraMemV2-430M/5B after 800B tokensOpen-Benchmark accuracy: +0.3 pointsNo clear effect
add balance loss (beta=0.001) (ablation)UltraMemV2 · UltraMemV2-430M/5Btraining loss and accuracy: helps at TopM=47, hurts at TopM=94Mixed
increase number of UltraMemV2 layers (2, 5, 10, 20) (ablation)UltraMemV2 · UltraMemV2-430M/5BOpen-Benchmark accuracy: +2.3, +0.9, +1.1Improved
put a memory layer in every transformer block (ablation)UltraMemV2 · UltraMemV2gain from continued training: +5.7 (4 layers) -> +8.3 (every block, 1.25B/12.5B)Improved
raise activated values per layer from top256 to top768 (ablation)UltraMemV2 · UltraMemV2-2.5BOpenbench All / Hardbench All (after 250B CT): 68.3 -> 69.1; 27.9 -> 30.0Improved
use Ring instead of NoRing value sharing (S9) (ablation)UltraMemV2 · UltraMemV2-500M/6.8BOpen-Benchmark accuracy: about +1 pointImproved
share value tables across layers (S9-Ring) (ablation)UltraMemV2 · UltraMemV2-500M/6.8BOpen-Benchmark accuracy: +1 point; S16-Ring no further gainImproved
constant 1x value learning rate instead of 4x decaying to 1x (ablation)UltraMemV2 · UltraMemV2-500M/6.8B at 1.4T tokensOpen-Benchmark accuracy: +0.4 points at 1.4T; behind earlyMixed
use UltraMemV2 instead of SeedMoE-2.5B/30BSeedMoE · 32K long-contextIn-context learning: 21.6 -> 29.5Improved

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