Research, in context.

2024 / ACM MobiCom 2024

FlexNN: Efficient and Adaptive DNN Inference on Memory-Constrained Edge Devices

First author (1/5)

Adaptive DNN inference under tight device-memory budgets.

• Memory Management
• Scheduling
• Adaptive Inference

Edge AI

2026 / ACM MobiSys 2026

Vec-LUT: Vector Table Lookup for Parallel Ultra-Low-Bit LLM Inference on Edge Devices

Co-first author (1/6)

Vector table lookup makes parallel ultra-low-bit LLM inference fast on edge CPUs.

• Low-bit Inference
• Kernel Optimization
• Matrix Multiplication

Large Language Models (LLMs) · Edge AI

2026 / IEEE Transactions on Mobile Computing (TMC)

DWI: Efficient On-Device LLM Decoding Through Dynamic Width Selection on Nested Models

Third author (3/6)

Dynamic width selection on nested models for efficient on-device LLM decoding.

• Adaptive Inference

Large Language Models (LLMs) · Edge AI

2025 / IEEE Transactions on Mobile Computing (TMC)

Squeezer: Efficient Multi-DNN Inference for Edge Video Analytics via Cross-Model Scheduling

Fifth author (5/8)

Cross-model scheduling for efficient multi-DNN edge video analytics.

• Scheduling

Edge AI · Video Analytics

2022 / ISCA 2022

DIMMining: Pruning-Efficient and Parallel Graph Mining on Near-Memory-Computing

Sixth author (6/9)

Pruning-efficient, parallel graph mining with near-memory computing.

• Graph Mining
• Near-memory Computing

2026 / ACM MobiSys Companion 2026

Building Efficient Inference Systems for Resource-Constrained Edge AI Deployment

Sole author (1/1)

A synthesis of inference systems for resource-constrained edge AI deployment.

• Efficient Inference
• Adaptive Inference
• Inference Runtimes

Edge AI

2026 / ArXiv preprint

OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism

First author (1/6)

One shared KV cache coordinates concurrent action and language generation in VLAs.

• KV Cache Management
• Scheduling
• Continuous Batching
• Inference Runtimes

Vision-Language-Action Models (VLAs) · Edge AI

2026 / ArXiv preprint

Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots

Fifth author (5/11)

A portable inference runtime for embodied models across heterogeneous robots.

• Inference Runtimes
• Heterogeneous Computing
• Kernel Optimization
• Closed-loop Control

Vision-Language-Action Models (VLAs) · World-Action Models (WAMs) · Edge AI

2026 / ArXiv preprint

ActProbe: Action-Space Probe for Early Failure Detection of Generative Robot Policies

Co-first author (2/11)

Action-space probes detect failures early in generative robot policies.

• Failure Detection

Vision-Language-Action Models (VLAs)

2026 / ArXiv preprint

Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation

Fifth author (5/11)

In-context causal learning for generalizable embodied manipulation.

• In-context Adaptation
• Interaction Memory
• Causal Representation Learning
• Policy Adaptation
• Test-time Scaling

Vision-Language-Action Models (VLAs) · World-Action Models (WAMs)

2026 / ArXiv preprint

Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence

Seventh author (7/15)

A closed-loop embodied harness for self-evolving physical intelligence.

• Self-improvement
• Failure Recovery
• Closed-loop Control
• Failure Detection
• Robot Simulation
• Scheduling
• Heterogeneous Computing
• Quantization

Embodied Agents

2026 / ArXiv preprint

EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents

Ninth author (9/15)

Skill-aware reflection helps embodied agents evolve through experience.

• Reflection
• Self-improvement

Embodied Agents

2024 / ArXiv preprint

Personal LLM Agents: Insights and Survey about the Capability, Efficiency and Security

Fourth author · efficiency section lead (4/25)

A survey of personal LLM agents, with a focus on capability, efficiency and security.

• Personal Agents
• Efficient Inference
• Security and Privacy

Personal Agents · Large Language Models (LLMs)

2024 / ArXiv preprint

ChainStream: An LLM-based Framework for Unified Synthetic Sensing

Sixth author (6/9)

An LLM-based framework that unifies synthetic sensing.

• Context Sensing
• Program Synthesis
• Stream Processing

Large Language Models (LLMs)

2025 / EMNLP 2025

An Empirical Study of LLM Reasoning Ability Under Strict Output Length Constraint

Sixth author (6/11)

An empirical study of LLM reasoning with strict output-length budgets.

• LLM Reasoning
• Test-time Scaling

Large Language Models (LLMs) · Edge AI

2026 / Open-source project

Cosmos Lite: Efficient inference for Cosmos 3 robot policies

Project maintainer (project)

Quantization, kernels and parallel rollouts bring world-action models to 24 GB GPUs.

• Low-bit Inference
• Kernel Optimization
• Inference Runtimes
• Robot Simulation
• Scheduling

World-Action Models (WAMs)

2025 / Merged upstream contribution

GPTQv2 format support in vLLM

Open-source contributor (contribution)

Correct zero-point handling for low-bit and asymmetric GPTQv2 inference in vLLM.

• Low-bit Inference
• Kernel Optimization

Large Language Models (LLMs)