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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add SinkTrack context anchoring over CoT | Qwen2.5-VL-7B-Instruct · M3CoT | accuracy: 44.11 -> 66.94 | Improved |
| add SinkTrack over CoT | Llama3.1-8B-Instruct · QuAC | accuracy: 46.95 -> 53.51 | Improved |
| add SinkTrack over CoT | Llama3.1-8B-Instruct · SQuAD2.0 | accuracy: 58.19 -> 79.83 | Improved |
| add SinkTrack over CoT | Qwen2.5-VL-7B-Instruct · POPE | accuracy: 83.65 -> 85.47 | Improved |
| add SinkTrack over Direct | Gemma3-12B-Instruct · RealWorldQA | accuracy: 57.34 -> 57.91 | No clear effect |
| add SinkTrack over Direct | MiniCPM3-4B · SQuAD2.0 | Macro-F1: 38.61 -> 36.83 | Hurt |
| use CoT prompting instead of Direct | Qwen2.5-7B-Instruct · QuAC | accuracy: 49.47 -> 36.48 | Hurt |
| replace hard injection by soft injection (hard injection replaces BOS Value vector) (ablation) | BOS context injection · text/multimodal | output quality | Hurt |
| inject every 5 layers instead of every layer (ablation) | soft injection · M3CoT (Qwen2.5-VL-7B-Instruct) | accuracy: 51.02 -> 60.04 | Improved |
| use decaying injection strength instead of increasing (ablation) | soft injection (intermittent) · M3CoT (Qwen2.5-VL-7B-Instruct) | accuracy: 59.87 -> 60.69 | Improved |
| soft injection with every 5 layers vs CoT baseline (ablation) | Qwen2.5-VL-7B-Instruct · M3CoT | accuracy: 43.83 -> 60.04 | Improved |
| drop mean-pooling and use the full uncompressed context in cross-attention (ablation) | SinkTrack · QuAC | accuracy | Improved |
Closest papers by meaning
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.93)
- Attention Sinks: A 'Catch, Tag, Release' Mechanism for Embeddings (NeurIPS 2025 · similarity 0.92)
- When Attention Sink Emerges in Language Models: An Empirical View (ICLR 2025 · similarity 0.91)
- On the Nature of Attention Sink that Shapes Decoding Strategy in Omni-LLMs (NeurIPS 2026 · similarity 0.91)
- Efficient Streaming Language Models with Attention Sinks (ICLR 2024 · similarity 0.90)
- Rope to Nope and Back Again: A New Hybrid Attention Strategy (NeurIPS 2025 · similarity 0.89)
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.
