SINCE 2025
Meta
Research Scientist
- Learned Indexes on Real Storage
- Globally Sorted, Locally Unsorted
- Storage Systems into LLM Inference
- Benchmarks & Honest Evaluation
Research Scientist · Meta · Univ. of Minnesota
Wenlong Wang
San Jose, California, United States
The systems researcher who made learned indexes work on storage that takes writes.
Wenlong Wang is a Research Scientist at Meta's MSL Infra, where he optimizes inference for media-generation models and contributes multi-modal and speculative-decoding paths to vLLM and LMCache. He earned his computer-science Ph.D. at the University of Minnesota under David Du, building TieredKV, an LSM-plus-Learned-Index key-value store published at ACM SYSTOR 2025. One thread runs through it all: indexing a huge, read-skewed key-value set fast — from SSD garbage collection to DNA test tubes to the LLM KV cache.
- 89.2kvLLM stars — multi-modal & spec-decoding contributor
- 4.32×TieredKV read throughput vs. LSM SOTA (SYSTOR'25)
- 11.2kLMCache KV-cache-layer stars
- 19×VL-DNA DNA-tube capacity gain (Blawat)
HAPPENING NOW
STANDING TODAY
SINCE 2025
vLLM
Open Source Contributor
HOW THEY THINK4 BELIEFS
Four design instincts run from SSD garbage collection to the LLM KV cache.
THE IMPACT
One method — index and serve a read-skewed key-value set fast — across papers and inference stacks.
EVERYTHING ELSE · 4 WORKS
THE TIMELINE
One question — how do you index and serve a huge, read-skewed key-value set fast — carried from SSDs and DNA to LLM caches.
The whole record, newest first – tap any moment and Aster tells the story.
2025–PRESENT
Optimizing LLM inference at Meta and in the open
In 2025 Wang carried his storage instinct into LLM inference. He contributes multi-modal support and speculative decoding to vLLM (89.2k stars), core-developed its JAX/TPU backend during a Google internship, and now works on media-generation inference at Meta's MSL Infra. The KV cache he tunes is the same read-skewed key-value working set his thesis modeled.
“Working on LLM inference optimization (multimodal support, spec decoding, RL support etc.)”
2015–2025
From HKUST to a Minnesota PhD in intelligent storage
Across a decade Wang followed one throughline from database storage structures at HKUST to a Minnesota Ph.D. under David Du in the CRIS lab. Two internships sharpened it from opposite ends — Alibaba (2021) inside the SSD's flash translation layer, Futurewei (2022) rebuilding the index above it with a 2.05× bucketed leaf-node win. The thesis fused them into LearnedKV/TieredKV, with BLI and VL-DNA alongside.
“While existing approaches use learned indexes primarily as auxiliary components within LSM trees, LearnedKV employs a two-tier design where the LSM tree handles recent write operations while a separate Learned Index accelerates read performance.”