Real life is where context gets hard

Tencent Hunyuan introduces CL-bench Life, a benchmark for evaluating context learning ability in real-life settings, containing 405 context-task pairs and 5,348 human-written rubrics.

Evaluations of 12 language models show they solve only 14.5% of tasks on average, with the best model (GPT-5.5 High) solving 22.2%.