Research map / Architectures and efficiency / RNNs and normalization
Towards Optimal Caching and Model Selection for Large Model Inference
Banghua Zhu, Ying Sheng, Lianmin Zheng, Clark Barrett, Michael Jordan, Jiantao Jiao
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 |
|---|---|---|---|
| replace LFU eviction with LEC (cost-aware) eviction | large-model-only caching · Lambada offline, alpha=0.8 | cumulative cost (FLOPs, x10^3): 10.81 -> 10.36 | Improved |
| add learned model selector to LEC caching | LEC+large · Lambada offline, alpha=0.8, acc 100% | cumulative cost (FLOPs, x10^3): 10.36 -> 2.50 | Improved |
| combine LEC with selector vs baseline LFU+large | LEC+selector · Lambada offline, alpha=0.8, acc 100% | cumulative cost (FLOPs, x10^3): 10.81 -> 2.50 | Improved |
| raise selector accuracy from 80% to 100% | LEC+selector · Lambada offline, alpha=0.8 | cumulative cost (FLOPs, x10^3): 4.76 -> 2.50 | Improved |
| use selector instead of cascade at 80% selector accuracy | LEC caching · Lambada offline, alpha=0.2 | cumulative cost (FLOPs, x10^3): 1.50 -> 2.00 | Hurt |
| use selector instead of cascade at 100% selector accuracy | LEC caching · Lambada offline, alpha=0.2 | cumulative cost (FLOPs, x10^3): 1.50 -> 0.99 | Improved |
| use cascade instead of always-large model | LFU caching · Lambada offline, alpha=0.8 | cumulative cost (FLOPs, x10^3): 10.81 -> 11.80 | Hurt |
| use cascade instead of always-large model | LEC caching · Lambada offline, alpha=0.8 | cumulative cost (FLOPs, x10^3): 10.36 -> 4.11 | Improved |
| replace LFU eviction with LEC eviction | selector-based caching · OpenAssistant online, alpha=0.8 | cumulative cost (latency, x10^3): 21.14 -> 15.49 | Improved |
| combine LEC with selector vs baseline LFU+large | LEC+selector · OpenAssistant online, alpha=0.8 | cumulative cost (latency, x10^3): 28.24 -> 15.49 | Improved |
| use cascade instead of always-large model | LFU caching · OpenAssistant online, alpha=0.2 | cumulative cost (latency, x10^3): 9.31 -> 13.88 | Hurt |
| use cascade instead of always-large model | LEC caching · OpenAssistant online | cumulative cost (latency, x10^3): 8.74 -> 8.82 (alpha=0.2); 26.07 -> 20.31 (alpha=0.8) | Mixed |
Closest papers by meaning
- QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024 · similarity 0.88)
- ThinK: Thinner Key Cache by Query-Driven Pruning (ICLR 2025 · similarity 0.88)
- PatchKV: Weight Space Compensation of KV Cache (NeurIPS 2026 · similarity 0.88)
- Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.88)
- Tail-Optimized Caching for LLM Inference (NeurIPS 2025 · similarity 0.87)
- Efficient Evaluation of LLM Performance with Statistical Guarantees (NeurIPS 2026 · similarity 0.87)
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.
