Wenhao Li
Wenhao Li is a doctoral candidate in the Master-Doctor Program in Artificial Intelligence at Shandong University, supervised by Prof. Yilong Yin, and since September 2025 a joint Ph.D. candidate at the Shenzhen Loop Area Institute (SLAI) under Prof. Liqiang Nie. His research focuses on multimodal learning, few-shot learning, and Vision-Language-Action models, with work published at conferences including NeurIPS, ICLR, and ACM MM. He is a co-author of VLA-ATTC, an adaptive test-time compute framework for VLA models accepted as a poster at ICML 2026.
“This paper solves one of the most pressing deployment problems in physical AI: VLA models fail at exactly the wrong moments — complex, ambiguous situations — and this framework cuts those failures in half without requiring retraining or sacrificing real-time control speeds.”
Source→“On the hardest benchmark task ('Both pots on stove'), PI0 only succeeds 40% of the time. VLA-ATTC brings that to 58% — and PI0.5 from 54% to 68% (Table 1).”
Source→“Rather than running expensive multi-candidate evaluation on every timestep (the approach that kills competing methods), VLA-ATTC uses a lightweight uncertainty detector: generate two action candidates, measure their divergence using Dynamic Time Warping (DTW), and only trigger deep deliberation when they disagree.”
Source→“The combination of Chang Xu and Shan You suggests affiliation with a major Chinese AI lab or university. Their focus on plug-and-play inference augmentation without base model modification is a pragmatic, deployment-oriented research direction.”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.