PAL: Program-aided Language Models

PAL: Program-aided Language Models

ArXiv:https://arxiv.org/pdf/2211.10435

GitHub:https://reasonwithpal.com/

一、动机

  • 大模型与Chain-of-Thought可以很好地将一些复杂的问题分解为若干个子问题并进行逐步推理;
  • 但是对于一些较为复杂的数学运算,活着涉及到交大的数字符号运算时,即便大模型对任务的规划和分解思考是正确的,但是依然会存在计算错误,这些错误通常大模型很难很好地解决;
  • 因此,本文思考将一些复杂的推理任务先用大模型进行分解,其次让大模型生成python代码并基于解释器来实现计算。通过引入代码,可以弥补由于计算错误所带来的问题。

This bridges an important gap in chain-of- thought-like methods, where reasoning chains can be correct but produce an incorrect answer.

二、方法

本文提出Program-Aided Language Model(PAL)。

相比于Chain-of-thought,每一个exemplar中包含一个推理路径,这个推理路径时融合了自然语言和python代码。且最终只提供完整的变成代码,不提供最终答案。大模型在该prompt的引导下对目标测试样本进行推理和代码生成,最终借助python解释器获得最终答案。

下图展示了一个同时含有自然语言和python代码的推理路径:

PAL方法与CoT的对比图如下所示:

Exemplar的构建

对于评测数据集中,如果现有的工作如果已经提供了exemplar,则直接使用,否则则随机采样3~6个标注样本作为exemplar。

推理路径中的代码函数名称也要与原始变量名保持一致,采用下划线分割的形式定义。

For example, a variable that describes the number of apples in the basket should have a name such as num apples in basket. This keeps the generated code linked to the entities in the question.

三、实验

数据集:

  • 数学运算:GSM8K、SVAMP、ASDIV、MAWPS;
  • 符号推理:BBH-Hard
  • 算法推理:BBH-Hard

GSM8K-Hard:

作者通过启发式更改数字的方式构建了一个新的数据集,并基于这个数据发现50%的情况下大模型虽然给出正确的推理思路但是由于交大的数字计算存在错误导致最终预测错误。

符号推理的prompt样例:

数学运算实验结果:

四、实现

Prompt

针对数学运算、符号推理、算法运算三种类型的任务分别设计了带有编程语言和自然语言的prompt。

python 复制代码
MATH_CHAT_BETA_PROMPT = '''
Let's use python to solve math problems. Here are three examples how to do it,
Q: Olivia has $23. She bought five bagels for $3 each. How much money does she have left?

def solution():

"""Olivia has 23. She bought five bagels for 3 each. How much money does she have left?"""

money_initial = 23

bagels = 5

bagel_cost = 3

money_spent = bagels * bagel_cost

money_left = money_initial - money_spent

result = money_left

return result

复制代码
Q: Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?

def solution():

"""Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?"""

golf_balls_initial = 58

golf_balls_lost_tuesday = 23

golf_balls_lost_wednesday = 2

golf_balls_left = golf_balls_initial - golf_balls_lost_tuesday - golf_balls_lost_wednesday

result = golf_balls_left

return result

复制代码
Q: There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?

def solution():

"""There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?"""

computers_initial = 9

computers_per_day = 5

num_days = 4 # 4 days between monday and thursday

computers_added = computers_per_day * num_days

computers_total = computers_initial + computers_added

result = computers_total

return result

复制代码
How about this question?
Q: {question}
'''.strip()
相关推荐
lljss20201 小时前
5. 神经网络的学习
人工智能·神经网络·学习
jie*1 小时前
小杰深度学习(fourteen)——视觉-经典神经网络——ResNet
人工智能·python·深度学习·神经网络·机器学习·tensorflow·lstm
闲看云起1 小时前
论文阅读《LIMA:Less Is More for Alignment》
论文阅读·人工智能·语言模型·自然语言处理
jie*1 小时前
小杰深度学习(sixteen)——视觉-经典神经网络——MobileNetV2
人工智能·python·深度学习·神经网络·tensorflow·numpy·matplotlib
TGITCIC1 小时前
有趣的机器学习-利用神经网络来模拟“古龙”写作风格的输出器
人工智能·深度学习·神经网络·ai大模型·模型训练·训练模型·手搓模型
whltaoin2 小时前
AI 超级智能体全栈项目阶段五:RAG 四大流程详解、最佳实践与调优(基于 Spring AI 实现)
java·人工智能·spring·rag·springai
Piink2 小时前
网络模型训练完整代码
人工智能·深度学习·机器学习
luoganttcc2 小时前
在 orin 上 安装了 miniconda 如何使用 orin 内置的 opencv
人工智能·opencv·计算机视觉
JinchuanMaster2 小时前
cv_bridge和openCV不兼容问题
人工智能·opencv·计算机视觉
心勤则明2 小时前
Spring AI 文档ETL实战:集成text-embedding-v4 与 Milvus
人工智能·spring·etl