Lanxiang Hu

I'm a final-year PhD candidate at UCSD, fortunately advised by Prof. Hao Zhang and Prof. Tajana Šimunić Rosing, working on efficient and reliable AI systems. I am also a research intern at NVIDIA Cosmos Lab.

Previously I was an intern at Snowflake AI research. Before joining UCSD, I have spent wonderful time working as a visiting research intern with Prof. Song Han. I completed my undergrad degree at UC Berkeley with majors in CS and Physics.

I design inference acceleration algorithms and systems that turn idle compute into speedups. I also build evaluation infrastructures for interactive AI workloads. Some of my projects include:

• Training and inference algorithms for efficient AI: JetSpec++/JetSpec (preprint), Jacobi Forcing (ICML'26), CLLMs (ICML'24), OSD (ICML'24)
• Evaluation infrastructures: WMEval (preprint), VideoScience-Bench (ECCV'26), LMGame-Bench (ICLR'26), GameArena (ICLR'25)
• Hardware-aware systems: Stratum (MICRO'25), PockEngine (MICRO'23)

Email  /  Scholar  /  Twitter  /  Linkedin  /  Github

News

• [2026/11] We are releasing JetSpec++, a causal parallel drafter with self refinement that drafts and refines the next block while the current one is verified. It runs up to 1.5x faster than DSpark, with higher throughput and lower per-user latency at moderate concurrency.
• [2026/06] We are introducing JetSpec, bringing up to 10-11x acceptance length with 1000+TPS on Qwen3-8B and Qwen3-30B-A3B using causal parallel tree drafting.
• [2026/05] I am excited to join NVIDIA Cosmos Lab for summer internship!
• [2026/05] I am honored to be awarded with the 2026 Qualcomm Innovation Fellowship.
• [2026/05] Our papers on Jacobi Forcing, d3LLM, and AMA-Bench are accepted to ICML 2026.
• [2026/04] I am honored to be selected as one of the 2026 ML and Systems Rising Stars.
• [2026/02] We are releasing VideoScience-Bench, a benchmark for evaluating video models' scientific reasoning and world modeling capabilities.
• [2026/01] Our papers on Lmgame Bench and Stronger-MAS are accepted to ICLR 2026.

Selected Publications

* denotes equal contribution. See Google Scholar for the full publication list.

Training and Inference Algorithms for Efficient AI

JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting

JetSpec++: Foresight Parallel Drafting with Self Refinements

JetSpec trains a causal parallel draft head that drafts a whole candidate tree in one forward pass, with up to 9.64x speedup on MATH-500. JetSpec++ adds self refinement and drafts the next block while the current one is verified, running up to 1.5x faster than DSpark with higher throughput and lower per-user latency at moderate concurrency.

Fast and Accurate Causal Parallel Decoding using Jacobi Forcing

Jacobi Forcing trains causal parallel decoders that are able to generate high-quality drafts conditioning on noisy context. On coding and math tasks, Jacobi Forcing achieves up to 4.5x speeudp with minimal to no performance loss.

Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs

We introduce an on-policy RL algorithm, AT-GRPO and a training system for multi-agent system training. Stronger-MAS boosts planning task accuracy from a 14.0–47.0% from single-agent RL baseline to 96.0–99.5%. It also improves reasoning performance with an average gain of 7.62% on coding and 17.93% on math.

TrimLLM: Progressive Layer Dropping for Domain-Specific LLMs

we introduce an algorithm to progressively prune MHA and MLP layers during domain-specific SFT to achieve up to 5.7x speedup and 60% less memory consumption in comparison with state-of-the-art model compression algorithms.

CLLMs: Consistency Large Language Models

We show LLMs can be trained to operate LLMs as highly efficient parallel decoders with 2.4x to 3.4x speedup across a variety of benchmarks.

Online Speculative Decoding

We introduce online speculative decoding algorithm (OSD) with improved responsiveness, speculation accuracy and compatibility with LLM serving systems.

Evaluation Infrastructures

Benchmarking Scientific Understanding and Reasoning for Video Generation using VideoScience-Bench

VideoScience-Bench is a benchmark to evaluate video generative world models not only as generators but also as reasoners, requiring their generations to demonstrate scientific understanding consistent with expected physical and chemical phenomena.

Lmgame-Bench: How Good are LLMs at Playing Games?

We introduce lmgame-bench that evaluates latest large models with games and addresses evaluation challenges by providing scaffolds. We present quantitative analysis of the relationship between model gaming performance and results on existing benchmarks.

GameArena: Evaluating LLM Reasoning through Live Computer Games

We design and build a incentivized dynamic benchmarks to evaluate AI reasoning abilities extending beyond math and coding.

Media Coverage

Snowflake AI Research TechCrunch Hacker News Daily Top 3

Academic Services

Conference reviewer: ICML, ICLR, NeurIPS, COLM, ACL, ECCV.

© 2026 Lanxiang Hu. Design and source code from Jon Barron's website, powered by Jekyll.