Pengcheng Wang

My research focuses on efficient and adaptive Machine Learning Systems (MLSys), spanning inference optimization and parameter-efficient fine-tuning. Through joint algorithm and system design, I develop model optimization methods and content-aware runtime scheduling strategies for Large Language Models (LLMs) and Vision-Language Models (VLMs). Across edge GPUs, server GPUs, and custom AI accelerators, I aim to maximize task performance while meeting latency and throughput requirements within memory, cost, and energy budgets.

News

Oct 2026	I successfully defended my Ph.D. thesis, Content- and Resource-Aware Adaptive Inference for 3D Perception and Multimodal Understanding.
Sep 2026	Our paper CORA: Per-Slice Coherent Orthogonal Rotation for SVD-based Low-Rank Adaptation has been accepted to NeurIPS 2026.
Jun 2026	Our paper Look Less, Think Faster: Joint Token-Compute Adaptation for Multimodal LLMs has been accepted to ECCV 2026.
Feb 2026	Our proposal, Edge-Constrained Vision-Language Model Inference for Autonomous Driving, has been accepted by the NVIDIA Academic Grant Program for Researchers. I applied as a Ph.D. student together with my advisor, Prof. Somali Chaterji.
Jun 2025	Our paper Agile3D: Adaptive Contention- and Content-Aware 3D Object Detection for Embedded GPUs has been accepted to MobiSys 2025.

Research Interests

Adaptive Machine Learning Systems
Content- and Resource-Aware Compute Adaptation and Scheduling
Efficient Vision and Multimodal Models

Education

Ph.D., Elmore Family School of ECE, Purdue University · West Lafayette · Aug 2019 - Oct 2026
Advisor: Prof. Somali Chaterji
M.S., Tongji University · Shanghai · 2014 - 2017 (Excellent Graduate)
Advisor: Prof. Meisong Tong
B.E., Tongji University · Shanghai · 2010 - 2014 (Excellent Graduate)
Advisor: Prof. Meisong Tong

Experience

Machine Learning Engineer Intern

EmbodyX · Fall 2025, Spring & Summer 2026

Built foundation models for robotic systems
Optimized VLM inference using token compression
Applied model compression for efficient deployment at scale

Software Engineer Intern, AI ToolChain

Sunlune · Spring & Summer 2025

Developed and validated kernel, runtime, and driver software frameworks for AI accelerators
Integrated kernels and optimized runtime workflows for efficient inference of Llama-family LLMs
Performed feature testing, performance tuning, and cross-platform debugging

Generative AI Model Intern

Sunlune · Summer & Fall 2024

Developed AI-enabled design flow for high-performance digital circuit design
Designed Reinforcement Learning (RL) models for circuit generation
Collaborated with IC design engineers to capture design experience with AI models

Selected Publications

NeurIPS
CORA: Per-Slice Coherent Orthogonal Rotation for SVD-based Low-Rank Adaptation

Pengcheng Wang, Ziran Liu, Wei Wang, and Wei Jiang

In Advances in Neural Information Processing Systems, 2026 (Accepted)

Abstract arXiv Bib PDF

CORA is a parameter-efficient fine-tuning method that adapts pretrained weights through coherent rotations of their SVD bases and diagonal spectrum adjustments. It improves commonsense reasoning and code generation performance over existing low-rank adaptation methods while using fewer trainable parameters.
@inproceedings{wang2026cora, title = {CORA: Per-Slice Coherent Orthogonal Rotation for SVD-based Low-Rank Adaptation}, author = {Wang, Pengcheng and Liu, Ziran and Wang, Wei and Jiang, Wei}, booktitle = {Advances in Neural Information Processing Systems}, year = {2026}, }
ECCV
Look Less, Think Faster: Joint Token-Compute Adaptation for Multimodal LLMs

Pengcheng Wang, Zhiquan Wang, Jayoung Lee, Zhuoyan Xu, Ran Xu, Saurabh Bagchi, Yin Li, and Somali Chaterji

In Proceedings of the European Conference on Computer Vision, 2026

Abstract Bib PDF

SmartVL jointly adapts both input visual tokens and LLM compute for each sample, balancing vision-side efficiency and reasoning-side depth under a global budget. A token controller selects informative visual tokens while a compute controller adjusts LLM execution, and both are trained to cooperate through a shared budget signal. This coordination yields a better accuracy-efficiency trade-off on MLLM benchmarks than optimizing token pruning or compute skipping in isolation.
@inproceedings{wang2026smartvl, title = {Look Less, Think Faster: Joint Token-Compute Adaptation for Multimodal LLMs}, author = {Wang, Pengcheng and Wang, Zhiquan and Lee, Jayoung and Xu, Zhuoyan and Xu, Ran and Bagchi, Saurabh and Li, Yin and Chaterji, Somali}, booktitle = {Proceedings of the European Conference on Computer Vision}, year = {2026}, }
MobiSys
Agile3D: Adaptive Contention- and Content-Aware 3D Object Detection for Embedded GPUs

Pengcheng Wang, Zhuoming Liu, Shayok Bagchi, Ran Xu, Saurabh Bagchi, Yin Li, and Somali Chaterji

In Proceedings of the 23rd ACM International Conference on Mobile Systems, Applications, and Services, 2025

Abstract Bib PDF Video Code Poster

Agile3D is an adaptive 3D object detection system for embedded GPUs. It integrates a cross-model multi-branch execution framework with a contention- and content-aware reinforcement learning controller to dynamically balance detection accuracy and inference latency under changing input scenes and hardware contention.
@inproceedings{wang2025agile3d, title = {Agile3D: Adaptive Contention- and Content-Aware 3D Object Detection for Embedded GPUs}, author = {Wang, Pengcheng and Liu, Zhuoming and Bagchi, Shayok and Xu, Ran and Bagchi, Saurabh and Li, Yin and Chaterji, Somali}, booktitle = {Proceedings of the 23rd ACM International Conference on Mobile Systems, Applications, and Services}, year = {2025}, publisher = {ACM}, }
TODAES
Virtuoso: Energy- and Latency-Aware Streamlining of Streaming Videos on SOCs

Jayoung Lee, Pengcheng Wang, Ran Xu, Sarthak Jain, Venkat Dasari, Noah Weston, Yin Li, Saurabh Bagchi, and Somali Chaterji

ACM Transactions on Design Automation of Electronic Systems, 2023

Abstract DOI Bib

Virtuoso is an efficient and adaptive video object detection system that jointly optimizes accuracy, energy efficiency, and latency on system-on-chips. It uses a multi-branch execution kernel and a lightweight runtime scheduler to select operating points across the accuracy-energy-latency space.
@article{lee2023virtuoso, title = {Virtuoso: Energy- and Latency-Aware Streamlining of Streaming Videos on SOCs}, author = {Lee, Jayoung and Wang, Pengcheng and Xu, Ran and Jain, Sarthak and Dasari, Venkat and Weston, Noah and Li, Yin and Bagchi, Saurabh and Chaterji, Somali}, journal = {ACM Transactions on Design Automation of Electronic Systems}, year = {2023}, publisher = {ACM}, doi = {10.1145/3564289}, }
EuroSys
LiteReconfig: Cost and Content Aware Reconfiguration of Video Object Detection Systems for Mobile GPUs

Ran Xu, Jayoung Lee, Pengcheng Wang, Saurabh Bagchi, Yin Li, and Somali Chaterji

In Proceedings of the European Conference on Computer Systems, 2022

Abstract DOI Bib

LiteReconfig is an efficient adaptive video object detection framework for mobile GPUs. It uses a cost-benefit analyzer and a content-aware accuracy prediction model to decide which features to evaluate and which execution branch to run at inference time.
@inproceedings{xu2022litereconfig, title = {LiteReconfig: Cost and Content Aware Reconfiguration of Video Object Detection Systems for Mobile GPUs}, author = {Xu, Ran and Lee, Jayoung and Wang, Pengcheng and Bagchi, Saurabh and Li, Yin and Chaterji, Somali}, booktitle = {Proceedings of the European Conference on Computer Systems}, year = {2022}, publisher = {ACM}, doi = {10.1145/3492321.3519577}, }
SenSys
ApproxDet: Content and Contention-Aware Approximate Object Detection for Mobiles

Ran Xu, Chen-lin Zhang, Pengcheng Wang, Jayoung Lee, Subrata Mitra, Somali Chaterji, Yin Li, and Saurabh Bagchi

In Proceedings of the ACM Conference on Embedded Networked Sensor Systems, 2020

Abstract DOI Bib

ApproxDet is an adaptive video object detection framework for mobile devices. It dynamically selects execution branches to meet accuracy-latency objectives under changing video content, compute contention, and user requirements.
@inproceedings{xu2020approxdet, title = {ApproxDet: Content and Contention-Aware Approximate Object Detection for Mobiles}, author = {Xu, Ran and Zhang, Chen-lin and Wang, Pengcheng and Lee, Jayoung and Mitra, Subrata and Chaterji, Somali and Li, Yin and Bagchi, Saurabh}, booktitle = {Proceedings of the ACM Conference on Embedded Networked Sensor Systems}, year = {2020}, publisher = {ACM}, doi = {10.1145/3384419.3431159}, }