Yueqi Wang

ML Engineer @Google/YouTube, PhD in neuroscience

[Yueqi Wang @yueqiw]

@fchollet What’s the human brain counterpart of the symbolic harness?

[Marktechpost AI @Marktechpost]

Google AI Introduces STATIC: A Sparse Matrix Framework Delivering 948x Faster Constrained Decoding for LLM Based Generative Retrieval

STATIC (Sparse Transition Matrix-Accelerated Trie Index for Constrained Decoding) addresses the hardware inefficiency of standard prefix trees in LLM-based generative retrieval by replacing pointer-chasing traversals with vectorized sparse matrix operations. By flattening trie structures into Compressed Sparse Row (CSR) matrices, the framework achieves O(1) I/O complexity, enabling hardware accelerators like TPUs and GPUs to enforce business logic without the typical latency bottlenecks associated with irregular memory access. Deployed at scale on YouTube, STATIC delivers a 948x speedup over CPU-offloaded tries with a negligible per-step overhead of 0.033 ms, directly increasing fresh video consumption by 5.1% and significantly improving cold-start recommendation performance.....

Full analysis: https://t.co/r12tCUfLJN

Paper: https://t.co/piB5r9tMFb

Code: https://t.co/P7QNBBvsfI

@YouTube @GoogleDeepMind @GoogleAI @GoogleResearch

[Yueqi Wang @yueqiw]

It's been a lot of fun working on this research, enabling extremely efficient constrained decoding for YouTube's LLM-based recommender system (https://t.co/dAqqLfuLxx).

It's also a great add-on to our PLUM generative retrieval framework (https://t.co/bgy193ayWe).

[Sumit @_reachsumit]

Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators

YouTube presents a constrained decoding method that flattens prefix trees into sparse CSR matrices.

📝 https://t.co/a99bstsiXK
👨🏽‍💻 https://t.co/P2d1fAfxmH

[Yueqi Wang @yueqiw]

@swyx @eugeneyan @aiDotEngineer YouTube and Google Deepmind also released a paper a week ago describing this work:
https://t.co/mpvMwhRkWX

PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations

[Yueqi Wang @yueqiw]

Our #recsys2025 case study on leveraging multimodal LLM to improve video recommendations, led by @BoyuanLong !

[Sumit @_reachsumit]

LLM-Powered Nuanced Video Attribute Annotation for Enhanced Recommendations

@BoyuanLong et al. at Google use LLMs as annotators to achieve nuanced content understanding at scale for video recommendations.

📝https://t.co/kzjPCyXtHr

[Yueqi Wang @yueqiw]

@shuchaobi It will also help us solve many puzzles in the universe - AGI/ASI agents on silicon hardware can travel/explore across the universe much more easily than human as they don’t need oxygen, water, etc for survival, and are not limited by a single lifespan.

[Yueqi Wang @yueqiw]

We explored incorporating negative user feedback into the training objective of sequential retrieval models, and demonstrated its benefits using both live experiments and a counterfactual simulation framework that measures the recommender's responsiveness to user feedback.

[Yueqi Wang @yueqiw]

Our recent work on "Learning from Negative User Feedback and Measuring Responsiveness for Sequential Recommenders" has been accepted at the #RecSys2023 Industry Track!

We'll be presenting this work at the @ACMRecSys conference in Singapore.
https://t.co/23sNjPTsAu

[Yueqi Wang @yueqiw]

Interactive visualization of the organoid single-cell RNAseq data is also available in an online browser: https://t.co/kCds0r4TjC

[Yueqi Wang @yueqiw]

Our paper is out on @NatureComms! We generated human telencephalic organoids from stem cell-derived single neural rosettes, and studied autism-associated SHANK3 deficiency using transcriptomic and electrophysiological analysis.
https://t.co/KtC8GVrYMI
and https://t.co/h7UInLcv82

[Shcheglovitov Lab @shcheglovitov]

Excited to share our latest research, spearheaded by incredible @yueqiw @UUNeurobiology @UtahNeuroscie, on human #stemcells derived telencephalic #brain #organoids and #autism gene SHANK3 published with @SpringerNature in @NatureComms today! Read here: https://t.co/7bwW3ip1Af

[Nature Communications @NatureComms]

@shcheglovitov https://t.co/qU7Vo2GUNb

[Organoid News @organoidscience]

Dr. @AlexShcheglovit's lab (@shcheglovitov) led by Dr. Yueqi Wang (@yueqiw) investigated human telencephalic #organoids from stem cell-derived single neural 🧠 rosettes and the hemizygous deletion of an autism associated gene, SHANK3.

Out in @NatureComms: https://t.co/0mjKMmvAlK

[Yueqi Wang @yueqiw]

@AlexShcheglovit @NatureComms Thanks for the great mentorship Alex!

[Yueqi Wang @yueqiw]

👀 Heterogeneity and molecular programming of progenitors for motor neurons and oligodendrocytes https://t.co/dL3aeqsS7t

[Yueqi Wang @yueqiw]

Our preprint using human cerebral organoids, scRNAseq, and electrophysiology to study autism-associated SHANK3 deficiency https://t.co/9QlN5r280o

[Shcheglovitov Lab @shcheglovitov]

New preprint out! Modeling autism-associated SHANK3 deficiency using human cortico-striatal organoids generated from single neural rosettes | bioRxiv https://t.co/VusOdAVvs6

[Yueqi Wang @yueqiw]

Our study on the synaptic deficits in human neurons with autism-associated genetic mutations #autism #synapse

[Shcheglovitov Lab @shcheglovitov]

Check out our new paper in Mol Psychiatry (https://t.co/dpaoBOaiDH)! SHANK3 hemizygosity disrupts AMPA synapses and neuronal morphology in human neurons transplanted in the mouse brain. Huge kudos to @SimoneChiola, @KandyNapan, @yueqiw! #SHANK3, #22q13, #autism, #synapses, #iPSC