Research map / Architectures and efficiency / Attention and transformers
EpiCache: Episodic KV Cache Management for Long-Term Conversation on Resource-Constrained Environments
Minsoo Kim, Arnav Kundu, Han-Byul Kim, Richa Dixit, Minsik Cho
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 |
|---|---|---|---|
| episodic KV compression with block prefill instead of KVzip | Qwen3-4B, RealTalk · RealTalk cross-episode queries, 8K budget | GPT score: 35.9 -> 50.8 | Improved |
| episodic clustering and episode-specific caches vs KVzip in block prefill | Qwen3-4B, RealTalk · Avg across query types, 8K budget | GPT score: 36.0 -> 49.8 | Improved |
| add sensitivity-aware layer-wise budget allocation (ablation) | EpiCache, Qwen3-4B · RealTalk, 8K budget | GPT score: 49.8 -> 53.9 (alpha=2) | Improved |
| make layer allocation too sharp (alpha 8 vs 2 to 4) (ablation) | EpiCache, Qwen3-4B · RealTalk, 8K budget | GPT score: 53.9 -> 49.8 | Hurt |
| segment conversation by utterances instead of words or tokens (ablation) | EpiCache, Qwen3-4B · RealTalk, 8K budget | GPT score: 47.5 -> 49.8 | Improved |
| use Qwen3-Embedding instead of LLM embedding layer (ablation) | EpiCache, Qwen3-4B · RealTalk, 8K budget | GPT score: 43.0 -> 49.8 | Improved |
| scale embedding encoder from 0.6B to 4B (ablation) | EpiCache, Qwen3-4B · RealTalk, 8K budget | GPT score: 49.8 -> 50.6 | No clear effect |
| increase number of episodes E from 2 to 8 (ablation) | EpiCache, Qwen3-4B · RealTalk, 8K budget | GPT score: 47.9 -> 51.3 | Improved |
| sensitivity profiling with 1 vs 100 calibration samples, various domains (ablation) | EpiCache layer-wise allocation · LoCoMo and RealTalk, Qwen3-8B, 6K budget | F1 / GPT score | No clear effect |
| vary K-Means initialization and generation sampling seeds (ablation) | EpiCache · LoCoMo and RealTalk, Qwen3-8B | score: std below 0.5 | No clear effect |
| apply KVzip post-prefill eviction in block prefill setting | KVzip, LLaMA-3.2-3B · LongConvQA | LongConvQA score | Hurt |
| use block prefill with episodic KV and 8K budget instead of full KV | LLaMA3.2-3B, 90K-token history · 300-turn conversation efficiency | peak memory (GB): 36.3 -> 9.6 | Improved |
Closest papers by meaning
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.90)
- Training Free Exponential Context Extension via Cascading KV Cache (ICLR 2025 · similarity 0.90)
- LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models (ICML 2025 · similarity 0.90)
- Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning (ICLR 2025 · similarity 0.89)
- ThinK: Thinner Key Cache by Query-Driven Pruning (ICLR 2025 · similarity 0.89)
- RazorAttention: Efficient KV Cache Compression Through Retrieval Heads (ICLR 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.
