Jiaming Xu

Agentic AI inference can be accelerated by 2-6x with history-free speculative decoding that eliminates KV cache overhead and boosts acceptance rates by up to 70%.

Exact recovery of high-dimensional Procrustes matching is achievable with a polynomial-time algorithm, even at constant correlation levels.

LIT-GS achieves unprecedented robustness in mapping under illumination changes by integrating LiDAR geometry, outperforming state-of-the-art methods in low-contrast environments.

Achieving millimeter-level precision in state estimation, FAST-LIVGO outperforms existing methods by effectively handling dynamic conditions and degeneracy in odometry.

SB-RF achieves leading performance in speech enhancement with a one-step generative approach that outperforms traditional multi-step methods.

Most MLLMs still struggle with reflective decision-making and self-evaluation in Chinese mobile GUI environments, hindering their reliability in real-world interactions.

GPT-4's mobile proactivity is so bad (7.4% success) that a fine-tuned Qwen2 model more than doubles its performance, revealing a critical gap in current MLLMs and a path to improvement.