Research map / Architectures and efficiency / Mixture of experts
SpecMD: A Comprehensive Study On Speculative Expert Prefetching
Duc N.M Hoang, Mohammad Samragh, AJAY KUMAR JAISWAL, 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 |
|---|---|---|---|
| replace eviction policy with Least-Stale (ablation) | Eliseev '23 (OLMoE-1B-7B) · TTFT, OLMoE-1B-7B, 5% cache | TTFT (ms): 3076.8 -> 2748.1 | Improved |
| replace eviction policy with Least-Stale (ablation) | Zhu et al. '25 (OLMoE-1B-7B) · TTFT, OLMoE-1B-7B, 5% cache | TTFT (ms): 2257.2 -> 1473.1 | Improved |
| replace eviction policy with Least-Stale (ablation) | HOBBIT '24 (OLMoE-1B-7B) · TTFT, OLMoE-1B-7B, 5% cache | TTFT (ms): 1590.1 -> 1309.2 | Improved |
| replace eviction policy with Least-Stale (ablation) | MixtureCache '25 (OLMoE-1B-7B) · TTFT, OLMoE-1B-7B, 5% cache | TTFT (ms): 2664.6 -> 2297.5 | Improved |
| replace eviction policy with Least-Stale (ablation) | Zhu et al. '25 (Mixtral-8x7B) · TTFT, Mixtral-8x7B, 5% cache | TTFT (ms): 1620.3 -> 1460.8 | Improved |
| replace eviction policy with Least-Stale (ablation) | Zhu et al. '25 (Phi-3.5-MoE) · TTFT, Phi-3.5-MoE, 5% cache | TTFT (ms): 1905.6 -> 1987.6 | Hurt |
| replace eviction policy with Least-Stale (ablation) | HOBBIT '24 (Phi-3.5-MoE) · TTFT, Phi-3.5-MoE, 5% cache | TTFT (ms): 3458.6 -> 1933.6 | Improved |
| replace eviction policy with Least-Stale (ablation) | Eliseev '23 (Qwen1.5-MoE-A2.7B) · TTFT, Qwen1.5-MoE-A2.7B, 5% cache | TTFT (ms): 3724.0 -> 3433.4 | Improved |
| use Least-Stale instead of LRU eviction (ablation) | SpecMD · collision misses, 5% cache | collision miss rate: 1.6-1.9% vs 4.5-12.6% | Improved |
| use Least-Stale instead of Score-Based eviction (ablation) | SpecMD · collision misses, 5% cache | collision miss rate: 1.6-1.9% vs 42.6-60.9% | Improved |
| use score-based prefetching instead of top-k (ablation) | SpecMD with Least-Stale · hit rate / sync overhead, 5% cache | hit rate | Improved |
| use substitution of missing experts with cached experts (ablation) | SpecMD · GSM8K, 5% cache | exact match: 0.691 -> 0.073 | Hurt |
Closest papers by meaning
- MoE-SpAc: Efficient MoE Inference Based on Speculative Activation Utility in Heterogeneous Edge Scenario (NeurIPS 2026 · similarity 0.89)
- SpeCache: Speculative Key-Value Caching for Efficient Generation of LLMs (ICML 2025 · similarity 0.86)
- Value-Aware Stochastic KV Cache Eviction for Reasoning Models (NeurIPS 2026 · similarity 0.86)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.85)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.85)
- Mixture of Lookup Experts (ICML 2025 · similarity 0.85)
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.
