AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

How Can Mamba Learn In Context with Outliers and Generalize Provably?

Hongkang Li, Songtao Lu, Xiaodong Cui, Pin-Yu Chen, Meng Wang

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Training dynamics of in-context learning

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 nonlinear gating (Mamba instead of linear attention) (ablation)one-layer linear Transformer · Synthetic ICL, alpha>1/2classification error: smaller than 0.01Improved
add nonlinear gating (Mamba instead of linear attention) (ablation)three-layer linear attention · Synthetic ICL (d=72, alpha=0.5), FQtest accuracy: 93.68 -> 99.73Improved
place outliers closest to the query instead of randomMamba · Synthetic ICL (d=72, alpha=0.5)test accuracy: 99.67 -> 82.73Hurt
place outliers closest to the query instead of randomlinear attention Transformer · Synthetic ICL (d=72, alpha=0.5)test accuracy: 94.12 -> 93.96No clear effect
add CQ data in second half of training (ablation)Mamba · Synthetic ICL, CQ placementICL accuracy: 82.73 -> 95.45Improved
use 2 heads instead of 1 (ablation)linear Transformer · Synthetic ICL, R placementtest accuracy: 94.12 -> 95.08Improved
use 4 heads instead of 1 (ablation)linear Transformer · Synthetic ICL, R placementtest accuracy: 94.12 -> 90.90Hurt
replace linear attention with softmax attention (ablation)three-layer Transformer · Synthetic ICL, CQ placementtest accuracy: 93.96 -> 99.28Improved
raise outlier fraction alpha from 0.4 to 0.7linear Transformer · Synthetic ICL, R placementtest accuracy: 94.12 -> 67.78Hurt
raise outlier fraction alpha from 0.4 to 0.7softmax Transformer · Synthetic ICL, R placementtest accuracy: 99.26 -> 95.58Hurt
keep 10 principal components instead of 5DistilBERT sentence vectors classifier · SST-2 PCAaccuracy: 74.08 -> 78.90Improved
use Mamba instead of linear Transformer with outlier phrase 'James Bond' (ablation)3-layer 2-head model · SST-2 ICL (alpha=0.75, FQ and R placement)ICL accuracyImproved

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