# SpecMD: A Comprehensive Study On Speculative Expert Prefetching

arXiv: https://arxiv.org/abs/2602.03921 · HTML: https://mlsanity.com/paper/2602.03921/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- Placed in Efficient MoE training and inference systems, one of 12 directions in Mixture-of-experts.
- First posted on arXiv 3 Feb 2026.

## Placed in

- Mixture-of-experts / [Efficient MoE training and inference systems](https://mlsanity.com/moe/efficient-moe-training-and-inference-systems/)

## Built on by (most cited first, 3 of 3)

- [Cache-Aware Joint Router Adaptation for Memory-Efficient MoE Inference](https://mlsanity.com/paper/2609.04895/) (2026)
- [VisMMOE: Exploiting Visual-Expert Affinity for Efficient Visual-Language MoE Offloading](https://mlsanity.com/paper/2605.05899/) (2026)
- [Reproducible Evaluation of MoE Expert Caching: Replay Semantics, Workload Contamination, and Operating Regimes](https://mlsanity.com/paper/2608.07911/) (2026)

## Abstract (arXiv)

Mixture-of-Experts (MoE) models enable sparse expert activation, meaning that only a subset of the model's parameters is used during each inference. However, to translate this sparsity into practical performance, an expert caching mechanism is required. Previous works have proposed hardware-centric caching policies, but how these various caching policies interact with each other and different hardware specification remains poorly understood. To address this gap, we develop SpecMD, a standardized framework for benchmarking ad-hoc cache policies on various hardware configurations. Using SpecMD, we perform an exhaustive benchmarking of several MoE caching strategies, reproducing and extending prior approaches in controlled settings with realistic constraints. Our experiments reveal that MoE expert access is not consistent with temporal locality assumptions (e.g LRU, LFU). Motivated by this observation, we propose Least-Stale, a novel eviction policy that exploits MoE's predictable expert access patterns to reduce collision misses by up to 85× over LRU. With such gains, we achieve over 88% hit rates with up to 34.7% Time-to-first-token (TTFT) reduction on OLMoE at only 5% or 0.6GB of VRAM cache capacity.