EvaLearn
收藏资源简介:
EvaLearn是一个用于评估大型语言模型(LLMs)在挑战性任务中的学习和效率能力的开创性基准。它包含648个挑战性问题,分为182个序列,每个序列专注于一个任务类型。与大多数现有的并行评估基准不同,EvaLearn要求模型依次解决这些问题,从而能够利用从先前解决方案中获得的经验。EvaLearn提供五个全面的自动化指标来评估模型并量化其学习能力和效率。EvaLearn对九个前沿模型进行了广泛的基准测试,并观察到了不同的性能配置。该数据集旨在解决当前LLMs在学习和适应特定任务的能力方面的评估不足问题。
EvaLearn is a pioneering benchmark for evaluating the learning and efficiency capabilities of large language models (LLMs) on challenging tasks. It comprises 648 challenging problems grouped into 182 sequences, with each sequence focusing on a single task type. Unlike most existing parallel evaluation benchmarks, EvaLearn requires models to solve these problems sequentially, enabling them to leverage experience gained from prior solutions. EvaLearn provides five comprehensive automated metrics for evaluating models and quantifying their learning capabilities and efficiency. EvaLearn has been extensively benchmarked on nine state-of-the-art models, with distinct performance profiles observed. This dataset aims to address the current gap in evaluating LLMs' capabilities for learning and adapting to specific tasks.

- 1EvaLearn: Quantifying the Learning Capability and Efficiency of LLMs via Sequential Problem Solving复旦大学, 南京大学 · 2025年



