Lanxiang Hu

I'm a PhD candidate at UCSD, fortunately advised by Prof. Hao Zhang and Prof. Tajana Šimunić Rosing. My research interest is in building efficient and reliable AI models and systems at scale.

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

Currently my work lie in efficient AI and AI system evaluation. My prior work spans the following areas:

LLM inference acceleration: Jacobi Forcing (arXiv), CLLMs (ICML'24), OSD (ICML'24)
LLM post-training: Stronger-MAS (ICLR'26), LongPack (ICLR'25), TrimLLM (ACL'25)
Hardware-algorithm codesign: Stratum (MICRO'25), PockEngine (MICRO'23)
LLM evaluation: LMGame-Bench (ICLR'26), GameArena (ICLR'25)
Video model evaluation: VideoScience-Bench (arXiv)

Email / Scholar / Twitter / Linkedin / Github

News

[2026/01] Our papers on Lmgame Bench and Stronger-MAS are accepted to ICLR 2026.
[2025/12] Our new method for training native causal parallel decoder -- Jacobi Forcing is released! Try our models for AR quality with 4.5x speedup.
[2025/10] Our MAS training framework PettingLLMs (Stronger-MAS) is released! Try it out if you are working on agents!
[2025/09] I wrapped up my internship at Snowflake AI research. Go Snowflakes!
[2025/06] Lmgame Bench now supports model evaluation with and without gaming harness support (agentic workflow).
[2025/04] Our latest Lmgame Bench leaderboard and GamingAgent are now live! An exciting milestone toward evaluation of multi-turn interactive LLM/VLM agents.
[2025/01] Our papers on Game Arena and LongPack are accepted to ICLR 2025.
[2024/06] I joined Snowflake AI research for summer internship.
[2024/05] Our papers on CLLMs and OSD are accepted to ICML 2024.
[2024/03] A new family of parallel decoders, CLLMs are released. Try our codebase for 2~3 times inference speedup!

Selected Publications

* denotes for equal contribution. † denotes for significant contribution.

	Lmgame-Bench: How Good are LLMs at Playing Games? Lanxiang Hu, Mingjia Huo, Yuxuan Zhang†, Haoyang Yu†, Eric P. Xing, Ion Stoica, Tajana Rosing, Haojian Jin, Hao Zhang ICLR, 2026 arxiv / code / website / 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.
	Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs Yujie Zhao, Lanxiang Hu, Yang Wang, Minmin Hou, Hao Zhang, Ke Ding, Jishen Zhao ICLR, 2026 arxiv / code / website / 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.
	Stratum: System-Hardware Co-Design with Tiered Monolithic 3D-Stackable DRAM for Efficient MoE Serving Yue Pan, Zihan Xia, Po-Kai Hsu, Lanxiang Hu, Hyungyo Kim, Janak Sharda, Minxuan Zhou, Nam Sung Kim, Shimeng Yu, Tajana Rosing, Mingu Kang MICRO, 2025 arxiv / We present Stratum, a system–hardware co-design that enables efficient MoE serving by combining Mono3D DRAM, near-memory processing, and GPUs with access-aware data placement, achieving up to 8.29x higher throughput and 7.66x better energy efficiency than GPU-only baselines.
	GameArena: Evaluating LLM Reasoning through Live Computer Games Lanxiang Hu, Qiyu Li, Anze Xie, Nan Jiang, Ion Stoica, Haojian Jin, Hao Zhang ICLR*, 2025 arxiv / code / website / We design and build a incentivized dynamic benchmarks to evaluate AI reasoning abilities extending beyond math and coding.
	Scaling Long Context Training Data by Long-Distance Referrals Yonghao Zhuang, Lanxiang Hu, Longfei Yun, Souvik Kundu, Zhengzhong Liu, Eric P. Xing, Hao Zhang ICLR, 2025 arxiv / We show long distance referral is important to long context training, and design data pipeline to scale up constructing such data.
	TrimLLM: Progressive Layer Dropping for Domain-Specific LLMs Lanxiang Hu, Tajana Rosing, Hao Zhang ACL, 2025 arxiv / code / 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 Siqi Kou, Lanxiang Hu, Zhezhi He, Zhijie Deng, Hao Zhang ICML, 2024 arxiv / code / website / 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 Xiaoxuan Liu, Lanxiang Hu, Peter Bailis, Ion Stoica, Zhijie Deng, Alvin Cheung, Hao Zhang ICML, 2024 arxiv / We introduce online speculative decoding algorithm (OSD) with improved responsiveness, speculation accuracy and compatibility with LLM serving systems.
	PockEngine: Sparse and Efficient Fine-tuning in a Pocket Ligeng Zhu, Lanxiang Hu, Ji Lin, Wei-Chen Wang, Wei-Ming Chen, Chuang Gan, Song Han 56th IEEE/ACM International Symposium on Microarchitecture (MICRO-56), 2023 arxiv / website / We introduce PockEngine: a tiny, sparse and efficient engine to enable fine-tuning on various edge devices through sparse backpropagation and compile-time optimizations.