Automatic Prompt Augmentation and Selection with Chain-of-Thought from Label Data

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "Automatic Prompt Augmentation and Selection with Chain-of-Thought from 
Label Data"

By

Mr. Ka Shun SHUM


Abstract:

Chain-of-Thought (CoT) prompting is a technique that encourages the Large 
Language Models (LLMs) to generate intermediate rationales by providing 
reasoning sub-steps inside each few-shot demonstration. Recent work indicates 
CoT has significantly advanced the reasoning ability of LLMs and achieved 
superior performance in complex reasoning tasks. Despite the great success, 
previous CoT studies depend on carefully designed human-annotated rationale 
chains to prompt the LLMs. Specifically, the CoT demonstrations are sensitive 
to exemplars order, rationale complexity, rationale diversity and written 
style, making it challenging to be applied in various new applications.

In this thesis, we delve deeply into the factors that affect the performance of 
CoT prompting and propose a new strategy, Automate-CoT, capable of bypassing 
human engineering of CoT and producing robust demonstrations by first 
automatically augmenting rationale chains. Then it prunes low-quality chains to 
construct a pool of augmented rationale chains based on the labels. Finally, it 
selects the optimal combination of rationale chains from the high-quality pool 
by employing a variance- reduced policy gradient to estimate the significance 
of each demonstration.

Experimental results validate the effectiveness of Automate-CoT, showing 
competitive averaged improvements in arithmetic reasoning (+2.7%), commonsense 
reasoning (+3.4%), symbolic reasoning (+3.2%), and non-reasoning tasks (+2.5%). 
The findings of Automate-CoT enable rapid adaptation of CoT techniques to new 
tasks without large human effort.


Date:                   Tuesday, 6 August 2024

Time:                   10:00am - 12:00noon

Venue:                  Room 5501
                        Lifts 25/26

Chairman:               Prof. Raymond WONG

Committee Members:      Dr. Junxian HE (Supervisor)
                        Dr. Yangqiu SONG