Research output
Publications
Conference papers, newest first. * marks equal contribution.
2026
DySL-VLA: Efficient Vision-Language-Action Model Inference via Dynamic-Static Layer-Skipping for Robot Manipulation
Dynamic-static layer-skipping for vision-language-action models: adaptive layer bypass based on task requirements and motion significance, with 85.7× fewer trainable parameters than full fine-tuning.
Faster-MoA: Low-Latency Tree-Structured MoA Serving with Early Exit and Agent-Aware Prefill-Decode Overlap
Accelerating Mixture-of-Agents serving with a tree-structured aggregation architecture, semantic-similarity-based early exit, and agent-aware prefill-decode overlap — 10× faster inference with only ±1% accuracy variation.
CREATE: Cross-Layer Resilience Characterization and Optimization for Efficient yet Reliable Embodied AI Systems
Characterizing the reliability of LLM-based embodied AI systems across the application, system, and circuit layers, with error detection and correction techniques for efficient yet reliable embodied AI.