Research map / Architectures and efficiency / Attention and transformers
IceCache: Memory-Efficient KV-cache Management for Long-Sequence LLMs
Yuzhen Mao, Qitong Wang, Martin Ester, Ke Li
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 |
|---|---|---|---|
| cluster semantically similar tokens into pages with a DCI-tree instead of original-order pages | IceCache vs Quest/ArkVale-style paging · LongBench, Llama-3.1-8B-Instruct, budget 256 | average score: 42.8 -> 49.0 (ArkVale vs IceCache) | Improved |
| raise IceCache budget from 64 to 256 (ablation) | IceCache · LongBench, Llama-3.1-8B-Instruct | average score: 47.8 -> 49.0 | Improved |
| raise IceCache budget from 64 to 256 (ablation) | IceCache · LongBench, Mistral-7B-Instruct-v0.2 | average score: 39.0 -> 41.7 | Improved |
| raise IceCache budget from 64 to 256 (ablation) | IceCache · LongBench, Qwen3-32B | average score: 42.2 -> 43.1 | Improved |
| raise IceCache budget from 64 to 256 (ablation) | IceCache · LongBench, LongChat-7B-v1.5 | average score: 33.9 -> 35.0 | Improved |
| reuse selected KV page indices across layers (ablation) | IceCache · LongBench, Llama-3.1-8B-Instruct, budget 256 | average score: 49.0 -> 47.7 | No clear effect |
| reuse selected KV page indices across layers (ablation) | IceCache · 36k sequence latency | time to second token (s): 7.7 -> 5.9 | Improved |
| use IceCache instead of PQCache | KV-cache offloading · GSM8K CoT, 10% budget, Mistral-7B | accuracy: 46.2 -> 47.4 | Improved |
| use IceCache instead of PQCache | KV-cache offloading · LongGenBench, Llama-3.1-8B, budget 256 | average accuracy: 0.273 -> 0.331 | Improved |
| use IceCache instead of PQCache for time to second token | KV-cache offloading · 36k sequence latency | time to second token (s): 13.3 -> 7.7 | Improved |
| use IceCache instead of eviction methods SnapKV and StreamingLLM | KV-cache management · 36k sequence latency | time to second token (s): 0.55 (SnapKV) and 0.13 (StreamingLLM) vs 7.7 | Hurt |
| use IceCache with budget 64 instead of PQCache at budget 256 | KV-cache offloading · LongBench, Llama-3.1-8B | average score: 47.3 -> 47.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ArkVale — ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.92)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.92)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.92)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.92)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.92)
- SpeCache: Speculative Key-Value Caching for Efficient Generation of LLMs (ICML 2025 · 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.
