What is CL-bench and Why Did Tencent Hunyuan Develop It?

The CL-bench (Contextual Learning benchmark) is a new evaluation tool formally released by the team led by Yao Shunyu, Chief AI Scientist at Tencent Hunyuan. Its core mission is highly specific: to rigorously test whether a large language model (LLM) can learn new knowledge provided solely within the context of the current prompt and correctly apply it to solve a task.

The Key Limitation Addressed by CL-bench

Current state-of-the-art (SOTA) models often rely heavily on their 'parameterized knowledge'—the vast static information compressed into their weights during pre-training. While this makes them excellent at recalling known facts, they frequently fail when a task requires integrating new, dynamic information presented only in the input context, similar to how humans learn on the fly.

CL-bench Goal and Current Performance

The primary goal of CL-bench is to force models to move beyond memorization. The benchmark consists of 500 complex contextual tasks where the solution necessitates learning novel information not present in the model's training data. Early tests indicated a significant deficiency in current models; for instance, even high-performing models like GPT-5.1 (high) only achieved a task success rate of 23.7% on these contextual learning tests. This highlights the urgent need for benchmarks like CL-bench to shift LLM optimization toward true contextual adaptation.