AI Paper Atlas

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

SinkTrack: Attention Sink based Context Anchoring for Large Language Models

Xu Liu, Guikun Chen, Wenguan Wang

ICLR 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
add SinkTrack context anchoring over CoTQwen2.5-VL-7B-Instruct · M3CoTaccuracy: 44.11 -> 66.94Improved
add SinkTrack over CoTLlama3.1-8B-Instruct · QuACaccuracy: 46.95 -> 53.51Improved
add SinkTrack over CoTLlama3.1-8B-Instruct · SQuAD2.0accuracy: 58.19 -> 79.83Improved
add SinkTrack over CoTQwen2.5-VL-7B-Instruct · POPEaccuracy: 83.65 -> 85.47Improved
add SinkTrack over DirectGemma3-12B-Instruct · RealWorldQAaccuracy: 57.34 -> 57.91No clear effect
add SinkTrack over DirectMiniCPM3-4B · SQuAD2.0Macro-F1: 38.61 -> 36.83Hurt
use CoT prompting instead of DirectQwen2.5-7B-Instruct · QuACaccuracy: 49.47 -> 36.48Hurt
replace hard injection by soft injection (hard injection replaces BOS Value vector) (ablation)BOS context injection · text/multimodaloutput qualityHurt
inject every 5 layers instead of every layer (ablation)soft injection · M3CoT (Qwen2.5-VL-7B-Instruct)accuracy: 51.02 -> 60.04Improved
use decaying injection strength instead of increasing (ablation)soft injection (intermittent) · M3CoT (Qwen2.5-VL-7B-Instruct)accuracy: 59.87 -> 60.69Improved
soft injection with every 5 layers vs CoT baseline (ablation)Qwen2.5-VL-7B-Instruct · M3CoTaccuracy: 43.83 -> 60.04Improved
drop mean-pooling and use the full uncompressed context in cross-attention (ablation)SinkTrack · QuACaccuracyImproved

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