Research map / Architectures and efficiency / Attention and transformers
IndexMem: Learned KV-Cache Eviction with Latent Memory for Long-Context LLM Inference
Xintong Yang, Hao Gu, Binxing Xu, Lujun Li, Bei Liu, Jiacheng Liu, Qiyuan Zhu, Yike Guo, Sirui Han
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 learned indexer plus latent memory instead of heuristic eviction scores | KV eviction on Qwen3-8B · RULER 4K, 90% eviction | accuracy: 37.9 -> 46.2 vs KeyDiff | Improved |
| use IndexMem rather than SnapKV at high eviction | Mistral-7B-v0.3 · RULER 16K, 90% eviction | accuracy: 21.3 -> 52.9 | Improved |
| IndexMem vs TOVA on a long-context setting with heavy eviction | Llama-3.1-8B-Instruct · RULER 16K, 90% eviction | accuracy: 59.2 -> 56.0 | Hurt |
| IndexMem vs TOVA at 50% eviction | Mistral-7B-v0.3 · RULER 16K, 50% eviction | accuracy: 70.1 -> 68.5 | Hurt |
| add latent memory on top of indexer-only (ablation) | IndexMem · LongBench, Llama-3.1-8B-Instruct | score | Improved |
| add latent memory to indexer-only for NIAH (ablation) | IndexMem · NIAH, 50% compression | retrieval robustness | Improved |
| plug latent memory into SnapKV (SnapKV+Memory) (ablation) | SnapKV · LongBench tasks | score | Improved |
| increase decoding KV budget Bmax (ablation) | all methods · AIME25 and Math500, Qwen3-8B | accuracy | Improved |
| moderate eviction (about 50%) on LongBench triviaqa (ablation) | IndexMem · LongBench triviaqa | score | Improved |
| raise compression to 90% (ablation) | IndexMem · LongBench trec | score | Hurt |
Closest papers by meaning
- QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024 · similarity 0.93)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.92)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.92)
- RazorAttention: Efficient KV Cache Compression Through Retrieval Heads (ICLR 2025 · similarity 0.92)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.92)
- ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction (NeurIPS 2024 · similarity 0.92)
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.
