Research map / Large language models / Retrieval and knowledge
SR-KI: Scalable and Real-Time Knowledge Integration into LLMs via Supervised Attention
Bohan Yu, Wei Huang, Kang Liu
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 supervised attention training and top-k KB selection (SR-KI) instead of KBLaM | KBLaM-style KV injection · KB size 10000 QA | reference ID accuracy: 0.0087 -> 0.7800 | Improved |
| use SR-KI instead of KBLaM at small KB size | KB size 100 · knowledge BERTScore F1 | BERTScore F1: 0.8725 -> 0.8547 | Hurt |
| reuse retrieval-layer top-k indices in subsequent layers (ablation) | SR-KI · KB size 40000 | reference ID accuracy: 0.3600 -> 0.6940 | Improved |
| reuse retrieval-layer top-k indices in subsequent layers (ablation) | SR-KI · KB size 1000 | reference ID accuracy: 0.8167 -> 0.9467 | Improved |
| add supervised attention loss at the retrieval layer | SR-KI · KB retrieval, KB size 1000 | Recall@100: 0.4375 -> 0.9975 | Improved |
| add supervised attention training | SR-KI · unanswerable QA refusal accuracy | refusal accuracy | Hurt |
| use SR-KI instead of ICL at 100 KBs | in-context learning · KB size 100 | reference ID accuracy: 0.6730 -> 0.9837 | Improved |
| use SR-KI instead of KBLaM on alias-perturbed queries | KBLaM · generalization, KB size 1000 | ID generalization accuracy: 0.6810 -> 0.8089 | Improved |
| scale KB size from 1000 to 40000 | SR-KI · QA reference ID accuracy | reference ID accuracy: 0.9467 -> 0.6940 | Hurt |
| choose layer 25 as retrieval layer by single-layer injection (ablation) | Qwen2.5-7B-Instruct · layer identification (100 samples) | ID accuracy and EM accuracy | Improved |
Closest papers by meaning
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.89)
- KVLink: Accelerating Large Language Models via Efficient KV Cache Reuse (NeurIPS 2025 · similarity 0.88)
- RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval (NeurIPS 2025 · similarity 0.88)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.88)
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.88)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.88)
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.
