Brief

Six research papers from professors at the Chinese University of Hong Kong, Shenzhen, have been accepted to ICML 2026, a top-tier machine learning conference. The accepted works span areas such as efficient routing for large language models, risk-aware reasoning, and advanced sequence modeling. Notably, one paper, RACER, introduces a novel routing framework to minimize inference costs while controlling risks, and another, B-PAC Reasoning, offers an online method for efficient reasoning with guaranteed performance loss control. A third paper, MIMOMamba, proposes a new architecture for sequence modeling that jointly models temporal and cross-channel interactions. AI

The Shenzhen Institute for Big Data Research has had four of its research papers accepted by ICML 2026, a top-tier international conference in machine learning. Two of the papers introduce novel optimization techniques for large language models: AdaMeZO, an Adam-style zeroth-order optimizer that reduces memory overhead during fine-tuning, and Romberg-ZOGE, a method for higher-order bias reduction in gradient estimation. Another paper presents SCOPE, a framework for efficient video reasoning that uses a cloud-edge collaborative approach to decompose user queries. The fourth paper, MIMOMamba, proposes a new state-space model that jointly models temporal dependencies and cross-channel interactions with linear efficiency. AI