Research map / Vision-language and multimodal / Human and embodied modeling
InfoFlow KV: Information-Flow-Aware KV Recomputation for Long Context
Xin Teng, Canyu Zhang, Shaoyi Zheng, Danyang Zhuo, Tianyi Zhou, Shenji Wan
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 |
|---|---|---|---|
| use GLOBAL RoPE geometry instead of HL--TP for token selection (ablation) | Our · 2WikiMQA (passage split) | F1: 0.4722 -> 0.5019 | Improved |
| use GLOBAL RoPE geometry instead of HL--HP (ablation) | Our · 2WikiMQA (passage split) | F1: 0.4455 -> 0.5019 | Improved |
| use GLOBAL RoPE geometry instead of TL--TP (ablation) | Our · NarrativeQA (passage split) | F1: 0.1923 -> 0.2288 | Improved |
| use GLOBAL RoPE geometry instead of HL--HP (ablation) | Our · HotpotQA (passage split) | F1: 0.5529 -> 0.5954 | Improved |
| place important chunks closer to query (ascending) instead of descending (ablation) | Our + Reorder · 2WikiMQA (passage split) | F1: 0.4607 -> 0.5058 | Improved |
| place important chunks closer to query (ascending) instead of descending (ablation) | Our + Reorder · MuSiQue (passage split) | F1: 0.2875 -> 0.3285 | Improved |
| place important chunks closer to query (ascending) instead of descending (ablation) | Our + Reorder · HotpotQA (passage split) | F1: 0.6054 -> 0.5972 | Hurt |
| add chunk reordering to Our | Our, Qwen3-14B · 2WikiMQA (passage split) | F1: 0.5019 -> 0.5058 | No clear effect |
| add chunk reordering to Our | Our, Qwen3-14B · MuSiQue (passage split) | F1: 0.3386 -> 0.3285 | Hurt |
| add chunk reordering to Our | Our, Qwen3-14B · 2WikiMQA (fixed chunk 2048) | F1: 0.5089 -> 0.4773 | Hurt |
| use norm-based selection instead of CacheBlend | Qwen3-14B recomputation · 2WikiMQA (passage split) | F1: 0.4330 -> 0.5019 | Improved |
| use token-level selection instead of BM25 passage-level recompute | Qwen3-14B recomputation · 2WikiMQA | F1: 0.1904 -> 0.5019 | Improved |
Closest papers by meaning
- Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning (ICLR 2025 · similarity 0.87)
- Semantic Integrity Matters: Benchmarking and Preserving High-Density Reasoning in KV Cache Compression (ICML 2026 · similarity 0.87)
- KVLink: Accelerating Large Language Models via Efficient KV Cache Reuse (NeurIPS 2025 · similarity 0.87)
- KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction (NeurIPS 2025 · similarity 0.87)
- Retrieval from Within: An Intrinsic Capability of Attention-Based Models (NeurIPS 2026 · similarity 0.87)
- RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding (ICML 2025 · similarity 0.86)
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.
