When context is everything, AI models still struggle in the real world: Tencent

South China Morning PostEN 1 min read 100% complete by Vincent ChowFebruary 4, 2026 at 12:01 PM

AI Summary

short article 1 min

Tencent researchers, in collaboration with Fudan University, argue that AI models need improved "context learning" to be truly useful in real-world environments. Their research, published Tuesday, highlights that current models often fail in subtle but significant ways due to a lack of contextual understanding. To test this, they developed CL-bench, a benchmark evaluating 19 leading models across nearly 1,900 tasks. This research comes as Tencent aims to strengthen its foundational AI model efforts, led by former OpenAI researcher Vinces Yao Shunyu, after internal restructuring. Tencent's Hunyuan models currently lag behind domestic competitors like DeepSeek, and its consumer AI app Yuanbao trails ByteDance's Doubao in user numbers. The focus on context learning aims to address these shortcomings and improve AI performance in dynamic, real-world situations.

Key Entities & Roles

Key Player Opposition Context

DeepSeek 40% Yuanbao 40% Doubao 40% ByteDance 40% Shenzhen 50% Hunyuan 50% Vinces Yao Shunyu 60% Fudan University 60% OpenAI ×270% Yao 70% Tencent ×590%

Keywords

context learning 100% ai models 90% real-world environments 80% tencent 70% cl-bench 60% benchmark 60% model design 50% artificial intelligence 50% openai 40%