Final Answer Token in CoT Demonstrations
A specific token, such as ####, can be used in Chain-of-Thought demonstrations to clearly separate the final answer from the preceding reasoning steps. This practice teaches the language model a consistent and recognizable format for presenting its final conclusion.
0
1
Tags
Ch.3 Prompting - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Step 1: Calculate Boris's Final Apple Count (Boris and Beck's Apples Problem)
An arithmetic word problem about Boris and Beck's apples is used as a demonstration within a prompt for a language model. The demonstration includes the problem statement, a sequence of intermediate reasoning steps, and the final answer. What is the primary purpose of including the 'intermediate reasoning steps' in this context?
Evaluating Chain-of-Thought Demonstrations
Calculation Annotation in CoT Demonstrations
Final Answer Token in CoT Demonstrations
An arithmetic word problem is used to demonstrate a step-by-step reasoning process. The problem is: 'Boris starts with 100 apples. Beck has 23 fewer apples than Boris. Boris then gives 10 apples to Beck. How many more apples does Boris have than Beck in the end?' Arrange the following reasoning steps into the correct logical order to solve the problem.
Learn After
A developer is creating few-shot demonstrations to teach a language model to solve word problems. They notice the model's outputs are often verbose and fail to clearly state the final numerical answer, even when the reasoning steps are correct. Review the following demonstration from their prompt:
Q: A grocery store had 50 cans of soup. They sold 15 on Monday and received a new shipment of 25. How many cans do they have now? A: The store started with 50 cans. They sold 15, so 50 - 15 = 35. Then they received 25 more, so 35 + 25 = 60. The store now has 60 cans.
Which of the following critiques best identifies the primary weakness in this demonstration that is likely causing the model's inconsistent output format?
Improving Model Output Consistency
Refining a CoT Prompt for Programmatic Extraction