evaluation-study

研究通过预注册的受控实验比较了声明式动画规格与自由HTML/JavaScript代码两种LLM生成解释的方案，发现规格虽更短，但由于推理开销增加并未更便宜、更快或更可靠，小模型的规格可靠性甚至显著更低。

The paper presents a systematic cross-model evaluation of how large language models interpret verbal probability expressions, finding they track human benchmarks with fidelity but exhibit biases, particularly for negative expressions, with implications for human-AI uncertainty communication.