Core Concepts
Chain-of-X methods extend the concept of Chain-of-Thought to enhance the capabilities of Large Language Models across diverse tasks and domains.
Abstract
This survey provides a comprehensive overview of Chain-of-X (CoX) methods, which build upon the success of Chain-of-Thought (CoT) prompting for Large Language Models (LLMs).
The authors first introduce the background of CoT and define the generalized concept of CoX, where the "X" can represent various components beyond just reasoning thoughts, such as augmentation, feedback, and even models.
The survey then categorizes existing CoX methods based on the taxonomy of nodes (i.e., the X in CoX):
Chain-of-Intermediates: Methods that decompose complex problems into manageable subtasks (problem decomposition) or accumulate relevant information and evidence (knowledge composition).
Chain-of-Augmentation: Methods that incorporate additional knowledge in the form of instructions, histories, retrievals, and other domain-specific enhancements.
Chain-of-Feedback: Methods that leverage external or self-generated feedback to refine and improve the model's outputs.
Chain-of-Models: Methods that leverage the specialized expertise of multiple LLMs in a sequential manner.
Furthermore, the survey categorizes the applications of CoX methods across various tasks, including multi-modal interaction, factuality and safety, multi-step reasoning, instruction following, LLMs as agents, and evaluation tools.
The survey concludes by discussing potential future directions, such as causal analysis on intermediates, reducing inference cost, knowledge distillation, and end-to-end fine-tuning, highlighting the versatility and potential of CoX methods in enhancing LLM capabilities.
Stats
"Chain-of-Thought has been a widely adopted prompting method, eliciting impressive reasoning abilities of Large Language Models (LLMs)."
"Extending beyond reasoning thoughts, recent CoX methods have constructed the chain with various components, such as Chain-of-Feedback, Chain-of-Instructions, Chain-of-Histories, etc."
"CoX methods have been applied to tackle challenges in diverse tasks involving LLMs beyond reasoning, including multi-modal interaction, hallucination reduction, planning with LLM-based agents, etc."
Quotes
"The essence of CoT lies in its strategy to tackle complex problems by breaking them down into manageable intermediate steps."
"We refer to the X in CoX as the 'node' of the chain structure. Beyond the thoughts in CoT prompts, the X in CoX can take various forms tailored to specific tasks, including intermediates, augmentation, feedback, and even models."
"CoX methods have been instrumental in enhancing the interplay between textual and visual data in vision-language models."