Global R&D spending keeps climbing, yet the pace of scientific discovery hasn't kept up. Experimental cycles remain long, interdisciplinary collaboration faces structural barriers, and research workflows stay fragmented across instruments, disciplines and institutions.

The panel — titled "AI for Science: Beyond the Concept" — framed this as a systems problem as much as a science one, treating it as an engineering question: how to remove scientists from repetitive tasks and embed AI across the entire pipeline, from hypothesis to validation.

Yu Li, Director of the State Key Laboratory of Membrane Biology and Professor at Tsinghua University's School of Life Sciences, opened the session by describing AI's shift from a supporting tool towards a core part of research infrastructure. The goal, he said, is not to replace scientists but to free them from repetitive work so they can focus on insight and creative thinking.

Working with Suzhou National Laboratory, SLAI has built Owl·AuraID, a multi-agent system that automates experimental workflow from sample preparation through to data analysis. Rather than connecting to instruments through APIs, its agents operate equipment the way a person would — observing screens, clicking buttons and reading data directly.

That approach lets the system link software agents, embodied scientific agents and instruments that were never built to communicate with each other. The platform now covers six types of precision instruments, and has cut the workload of crystal structure analysis by 50.6%, reduced morphological analysis time from nine minutes to 7.5 minutes, and lifted AI autonomous completion rates from 33% to 80%.

BAAI's contribution targets a different constraint: the lack of a shared data format across neuroscience recordings. Its Wujie·Brainμ1.0 model, described as the world's first multimodal neuroscience foundation model, unifies EEG, calcium imaging and neural probe signals within a single encoding framework.

A study supported by the model, published in Science in June 2026, showed for the first time that memory reactivation regulates sleep in both directions — positive memories improving sleep quality, negative ones deepening fragmentation — with implications for treating sleep disorders linked to depression and anxiety. The model was trained on more than 70,000 nights of sleep data and has run over 12 months of automated analysis across partner laboratories.

CAS presented ScienceOne Omni, built on a three-layer architecture combining unified scientific data encoding, real-world knowledge alignment and domain-specific task decoding. Drawing on 170 million scientific publications and more than 8,000 specialised research tools, it operates across mathematics, physics, materials science, astronomy and other fields rather than being confined to one.

Within CAS, the model has compressed literature review from weeks to 20 minutes and lifted report generation efficiency by five to 10 times, and has been deployed across more than 100 research scenarios. Tested across more than 60 benchmarks, it outperformed general-purpose systems including Gemini and GPT on tasks such as chemical property prediction and protein binding site prediction. Applied to catalyst discovery with the Shanghai Institute of Ceramics, it cut design time from several months to 30 minutes and identified a candidate with 38% higher activity than existing options.

Key Takeaways

• →Four leading Chinese research institutions showed AI moving from a supporting tool to core research infrastructure at WAIC 2026.
• →SLAI's Owl·AuraID system operates lab instruments the way a human would, cutting crystal structure analysis workload by 50.6% and lifting autonomous completion from 33% to 80%.
• →BAAI's Wujie·Brainμ1.0 unifies previously incompatible neuroscience recording types, enabling a landmark Science study on memory and sleep regulation.
• →CAS's ScienceOne Omni spans multiple scientific disciplines with one model, outperforming Gemini and GPT on several specialised benchmarks.
• →ScienceOne Omni cut catalyst discovery design time from months to 30 minutes, identifying a candidate 38% more active than existing options.
• →Huawei's chips and joint engineering work with SLAI highlight a growing coupling between hardware and AI software stacks in research infrastructure.