AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

A Narrowing Geometry in Contaminated Reasoning

Jiakuan Xie, Pengfei Cao, Kang Liu, Jun Zhao

ICML 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes

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
apply Top-Subspace Suppression to contaminated modelQwen2.5-1.5B-Instruct contaminated · GSM8K restorationaverage consistency: 64.76 (LNE-Blocking) -> 73.01Improved
apply Top-Subspace Suppression to contaminated modelQwen2.5-7B-Instruct contaminated · GSM8K restorationaverage consistency: 60.29 (LNE-Blocking) -> 71.32Improved
use TSS instead of SamplingQwen2.5-7B-Instruct contaminated · GSM8K restorationrestoration rate: 9.09 -> 45.45Improved
use LNE-Blocking or TED baselines instead of TSScontaminated models · GSM8K retain setretain consistency: 97.19 -> 88.76 (7B, LNE-B.)Hurt
flatten activation eigenspectrum (tau=0.5) (ablation)Qwen2.5-7B-Instruct · relative truncation errorRTEMixed
sharpen activation eigenspectrum (tau=-0.5) (ablation)Qwen2.5-7B-Instruct · relative truncation errorRTEMixed
contaminate model by fine-tuning on leaked GSM8KQwen2.5 models · GSM8K leaked vs clean inputsmutual information (activations, gradients)Hurt
contaminate model by fine-tuning on leaked GSM8KQwen2.5 models · activation covarianceECIHurt

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