When context is everything, AI models still struggle in the real world: Tencent
Despite rapid advances, today’s top AI models are still “brittle” in messy, real-world environments, according to new Tencent research
AI developers need to place “context learning” at the centre of future model design if their products are to become genuinely useful outside controlled environments, according to researchers from Tencent and Fudan University’s Institute of Trustworthy Embodied AI.
The research comes as Yao – a former star researcher at OpenAI – seeks to reinvigorate Tencent’s foundational model efforts after a series of internal restructurings.
The Shenzhen-based conglomerate’s Hunyuan models trail domestic rivals such as DeepSeek while its flagship consumer AI app Yuanbao had roughly half the number of users of ByteDance’s market-leading Doubao as of January.
To test levels of context learning ability among existing models, Tencent’s researchers developed a new benchmark called CL-bench, testing 19 leading models across 1,899 tasks designed to measure on-the-job learning.