arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2306.04441cs.CL
收藏

STEPS:一个用于序列任务中顺序推理的基准

STEPS: A Benchmark for Order Reasoning in Sequential Tasks

University of California, Santa Barbara(加州大学圣巴巴拉分校)

Weizhi Wang, Hong Wang, Xifeng Yan

AI总结:

本文提出 STEPS 基准,通过食谱下一步合理性判断和多项选择两个子任务评估 LLM 的动作顺序推理能力,发现提示方法明显落后于微调方法。

摘要:

各种人类活动都可以被抽象为自然文本中的动作序列,例如烹饪、维修、制造等。此类动作序列高度依赖执行顺序,而动作序列中的顺序混乱会导致机器人或 AI 智能体后续任务执行失败。因此,为验证当前神经模型在序列任务中的顺序推理能力,我们提出了一个具有挑战性的基准,名为 STEPS。STEPS 包含两个子任务设置,分别侧重于判断食谱中给定下一步的合理性,以及从多项选择题中选择合理的步骤。我们描述了数据构建和任务形式化,并对大多数重要的大型语言模型(LLM)进行了基准测试。实验结果表明:1)对于 LLM 而言,通过零样本提示或少样本上下文学习来解决序列任务中动作顺序的常识推理颇具挑战;2)在 STEPS 上,提示方法仍显著落后于基于微调的方法。

英文摘要:

Various human activities can be abstracted into a sequence of actions in natural text, i.e. cooking, repairing, manufacturing, etc. Such action sequences heavily depend on the executing order, while disorder in action sequences leads to failure of further task execution by robots or AI agents. Therefore, to verify the order reasoning capability of current neural models in sequential tasks, we propose a challenging benchmark , named STEPS. STEPS involves two subtask settings, focusing on determining the rationality of given next step in recipes and selecting the reasonable step from the multi-choice question, respectively. We describe the data construction and task formulations, and benchmark most of significant Large Language Models (LLMs). The experimental results demonstrate 1) The commonsense reasoning of action orders in sequential tasks are challenging to resolve via zero-shot prompting or few-shot in-context learning for LLMs; 2) Prompting method still significantly lags behind tuning-based method on STEPS.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑