PhD student in Electrical and Computer Engineering
Carnegie Mellon University
Email:
I am a PhD student in Electrical and Computer Engineering at Carnegie Mellon University, advised by Prof. Beidi Chen. Previously, I worked with Prof. Diana Marculescu on edge AI and Prof. Kurt Keutzer on model quantization. I received my bachelor's degree in Electrical and Computer Engineering from The University of Texas at Austin.
My research focuses on efficient post-training and inference for large language models, with an emphasis on reasoning, long-context generation, and reinforcement learning. I welcome conversations and collaborations in these areas.
Efficient training, evaluation, and inference for reasoning models.
Efficient Reinforcement Learning explores practical training when efficient rollout actors differ substantially from the policy being trained, including budgeted rejection sampling and selective rollouts in Jackpot, GRESO, and Sparrow.
Long-context Reasoning scales up synthetic data generation to study how language models behave as context length and reasoning complexity increase, with controllable evaluation across both dimensions through GSM-Infinite, and MagicPIG.
Efficient LLM Inference reduces the computation and memory required for language-model inference while preserving capability, spanning contextual sparsity and system-level analysis in Vortex, Sirius, LLM Inference Unveiled, Kinetics, HexGen, and DQRM.
(Early in Undergraduate) Edge AI studies efficient deployment and training on resource-constrained systems through Play It Cool, and ANT.
* denotes equal contribution.
Preprint, 2026
A dynamic sparse-rollout schedule balances actor-policy mismatch and rollout speed for stable long-context reinforcement learning. An intermediate version is accepted in SPOT at ICLR, 2026.
ICLR, 2026
Budgeted rejection sampling enables efficient RL training under extreme actor-policy mismatch.
Preprint, 2026
A programmable system for rapidly expressing, deploying, and evaluating sparse-attention algorithms in modern LLM serving stacks.
NeurIPS, 2025 TTODLer-FM at ICML, 2025 Best Paper Award TTODLer-FM at ICML, 2025 Oral
A test-time scaling law that accounts for both computation and memory-access costs and motivates sparse attention for efficient scaling.
NeurIPS, 2025
GRESO predicts and skips uninformative prompts before rollout, reducing reinforcement learning overhead.
ICML, 2025 LCFM at ICML, 2025
A controllable benchmark for evaluating language models under increasing context length and reasoning complexity.
ICLR, 2025 Spotlight
An LSH-based heterogeneous system that samples attention with theoretical guarantees for efficient long-context generation.
NeurIPS, 2024
A lightweight correction method recovers reasoning accuracy in contextually sparse language models while retaining efficiency.
Preprint, 2024
A survey of algorithm, system, and hardware techniques for LLM inference, accompanied by a roofline modeling tool for analyzing deployment bottlenecks. Cited over **200 times** (2026.07).
ICML, 2024
A distributed inference engine for deploying foundation models across heterogeneous GPUs and network connections.
Preprint, 2023
A 4-bit quantized recommendation model and sparse-gradient training system that reduce model size and communication overhead without sacrificing accuracy.
DyNN at ICML, 2022 Oral
Dynamic switching between efficient and powerful networks prevents thermal throttling during sustained inference on edge devices.
CVPRW ECV, 2022
ANT exploits semantic redundancy between video frames by adapting a purpose-fit network over time.
Reviewer for ECCV 2023, MLSys 2024, and NeurIPS 2025. Selected as a Top Reviewer at NeurIPS 2025, free trip ✈️ yayyy!