Efficiency across the AI stack
Modern AI is powerful but expensive. Our mission is to close the gap between what models can do and what real devices can afford — co-designing algorithms and systems so that intelligence becomes fast, compact, and sustainable.
Efficient LLMs & Multimodal AI
Large language and vision-language models deliver remarkable capability at a steep compute and memory cost. We design token-pruning, efficient-attention, and serving techniques that cut this cost while preserving quality — including training-free methods that apply to already-deployed models. Recent work includes ZOO-Prune (training-free token pruning) and VisRef (visual refocusing for test-time scaling).
Scalable Agentic AI
Agentic and reasoning systems achieve strong performance by scaling test-time computation — but at a rapidly growing token and inference cost. We study how to make long-horizon LLM agents affordable: benchmarking their capabilities, understanding agent harness design, and reducing the tokens spent per task. Recent work includes SkillRet (a large-scale benchmark for skill retrieval in LLM agents) and Agent Harness Engineering: A Survey.
Brain-Inspired & Neuromorphic Computing
Spiking neural networks perform event-driven, sparse computation that maps naturally onto energy-efficient neuromorphic hardware. We study training methods, coding schemes, distillation, and ANN-to-SNN conversion — and pair them with in-memory computing — to bring this efficiency to modern deep learning.
Continual & Adaptive Learning
Deployed models face shifting, open-world data. We develop prompt-based continual learning and backpropagation-free test-time adaptation so models keep improving after deployment. Recent work includes Open-World Dynamic Prompt learning and Probabilistic Gaussian Alignment for backpropagation-free test-time adaptation.
Hardware-Aware Co-Design
Efficiency only matters if it maps to real accelerators. We co-design algorithms with hardware — dataflows, crossbar-based and in-memory computing, and transform reuse — so that compact models translate into measurable speedups and energy savings end to end.
Model Compression
We shrink neural networks so they fit real hardware budgets: post-training and low-data quantization, structured and unstructured pruning, and neural architecture search that preserve accuracy for edge and on-device deployment. Representative work includes GenQ (quantization in low-data regimes) and our neural-architecture-search and lottery-ticket studies.