Xi Liu

CIVIC is a path-consistent compact visual inference framework that achieves genuine hardware efficiency in Vision-Language Models by maintaining contiguous sequence representations across all inference stages, significantly reducing memory and latency.

The paper proposes a pose-conditioned, permutation-equivariant denoiser to accurately reconstruct work zone geometry using noisy Ultra-Wideband (UWB) range data from connected and autonomous vehicles (CAVs).

This paper proposes Learning to Allocate (L2A), an end-to-end framework for resource-adaptive inference in Large Language Models (LLMs) using budget-conditioned and input-aware gating networks.

This paper introduces Bifocal dLLMs (R2LM), a new paradigm for discrete diffusion language models that combines causal and bidirectional attention for improved throughput and generation quality.

This paper proposes Diffusion-GR2, a method to convert an autoregressive reasoning re-ranker into a block-diffusion re-ranker while maintaining accuracy and increasing speed.

This paper proposes SCOReD, a framework for optimizing chain-of-thought (CoT) distillation in the recommendation domain by parsing teacher traces into typed segments, scoring their importance, and dynamically selecting edits based on the student's output distribution.

This paper introduces Audio-Dependency Filtering (ADF) pipeline for Audio-Dependent Question Answering (ADQA) task in DCASE~2026, achieving top overall and sub-10B accuracy.

This paper introduces DGNA, a methodology to unveil the Non-Uniform Memory Access (NUMA) architecture of GPU memory hierarchy through microbenchmarking and data analysis.