AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Transformers with Endogenous In-Context Learning: Bias Characterization and Mitigation

Haotian Wang, Hao Zou, Haoxuan Li, Haoang Chi, Yang Shi, Yuanxing Zhang, Wenjing Yang, Xinwang Liu, Zhouchen Lin

ICLR 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
Increase DDbias ICL examples from 15 to 60DDbias on LLaMA · Yelp RPIMAE: 22.8 -> 16.7Improved
Increase DDbias ICL examples from 15 to 60DDbias on LLaMA · Yelp RORMAE: 0.36 -> 0.16Improved
Apply DDbias with 60 ICL examples instead of vanilla LLaMALLaMA · Yelp RPIMAE: 23.8 -> 16.7Improved
Apply DDbias with 60 ICL examples instead of vanilla LLaMALLaMA · Yelp RORMAE: 0.46 -> 0.16Improved
Apply DDbias with only 15 ICL examples versus DMCEE baselineLLaMA · Yelp RPIMAE: 21.2 -> 22.8Hurt
Apply DDbias with 15 ICL examples versus SC baselineLLaMA · Yelp RORMAE: 0.24 -> 0.36Hurt
Apply DDbias to 5-layer non-linear transformerTransformer L=5 · Polynomial f, Conf@2.0prediction error: 0.600 -> 0.260Improved
Apply DDbias to 5-layer non-linear transformerTransformer L=5 · Polynomial f, Conf@0.5prediction error: 0.280 -> 0.115Improved
Apply DDbias to 7-layer non-linear transformerTransformer L=7 · Polynomial f, Conf@2.0prediction error: 0.750 -> 0.290Improved
Apply DDbias to 7-layer non-linear transformerTransformer L=7 · Polynomial f, Conf@0.5prediction error: 0.320 -> 0.135Improved
Increase confounding strength from Conf@0.5 to Conf@2.0 (biased TF) (ablation)Vanilla TF L=5 · Polynomial fprediction error: 0.280 -> 0.600Hurt
Increase confounding strength from Conf@0.5 to Conf@2.0 (DDbias) (ablation)DDbias L=5 · Polynomial fprediction error: 0.115 -> 0.260Hurt

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