AI Paper Atlas

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

Attention Sinks as Internal Signals for Hallucination Detection in Large Language Models

Jakub Binkowski, Kamil Adamczewski, Tomasz Jan Kajdanowicz

ICML 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering · 0 citations (OpenAlex)

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 SinkProbe sink-score features instead of prior attention-based detectorsSinkProbe · 7 hallucination benchmarks x 4 LLMsROC-AUC: best in 23 of 28 model-dataset pairsImproved
SinkProbe vs best baseline on UMWP with Llama3.2-3BSinkProbe · UMWP, Llama3.2-3BROC-AUC: 0.809 AttnLogDet vs 0.757Hurt
SinkProbe vs best baseline on GSM8K with Mistral-NemoSinkProbe · GSM8K, Mistral-NemoROC-AUC: 0.841 LookbackLens vs 0.811Hurt
SinkProbe vs best baseline on TruthfulQA with Phi3.5SinkProbe · TruthfulQA, Phi3.5ROC-AUC: 0.797 MTopDiv vs 0.761Hurt
SinkProbe vs best baseline on UMWP with Llama3.1-8BSinkProbe · UMWP, Llama3.1-8BROC-AUC: 0.838 -> 0.879Improved
increase top-k sink scores per head from 1 to between 2 and 5 (ablation)SinkProbe · GSM8K, UMWP, NQ-Open, TruthfulQA (Llama3.2-3B)ROC-AUCImproved
increase k beyond small values (k>=10) (ablation)SinkProbe · GSM8K, UMWP, NQ-Open, TruthfulQA (Llama3.2-3B)ROC-AUCNo clear effect
use sink scores instead of raw attention eigenvalues for top-k featuresSinkProbe · GSM8K, UMWP, NQ-Open, TruthfulQA (Llama3.2-3B)ROC-AUCImproved
train probe with L1 regularization (ablation)SinkProbe · GSM8K, NQ-Open, TruthfulQA, UMWPROC-AUC: retains 1% to 4% of featuresNo clear effect

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