Core Concepts
Introducing a method for supplementing large language models with argumentative reasoning to enable explainable and contestable decision-making.
Abstract
The paper introduces a novel framework called "Argumentative LLMs" that leverages large language models (LLMs) to construct argumentation frameworks, which then serve as the basis for formal reasoning in decision-making tasks. The key highlights are:
Argument Generation: The LLMs are used to generate arguments supporting and attacking a given claim. This results in an argumentation framework, which can be a tree of varying depth.
Argument Strength Attribution: The LLMs are also used to assign intrinsic strengths to the generated arguments, capturing their relative importance.
Argument Semantics: The argumentation framework is then evaluated using formal argument semantics, such as DF-QuAD or QEM, to determine the final decision.
The authors demonstrate the effectiveness of Argumentative LLMs experimentally in the decision-making task of claim verification, where they obtain results that are competitive with, and in some cases surpass, comparable state-of-the-art techniques. Importantly, the interpretable nature of the argumentation frameworks and formal reasoning means that any decision made by the supplemented LLM may be naturally explained to, and contested by, humans.
The authors also discuss the broader benefits of their approach, such as the ability to handle highly complex, uncertain, and high-stakes scenarios, the inherent calculation of uncertainty, and the potential for human-computer collaboration.
Stats
"The diversity of knowledge encoded in large language models (LLMs) and their ability to apply this knowledge zero-shot in a range of settings makes them a promising candidate for use in decision-making."
"However, they are currently limited by their inability to reliably provide outputs which are explainable and contestable."
Quotes
"Can the reasoning abilities of LLMs improve if they are made to argue with themselves?"
"Rather than prompting an LLM to produce 'thoughts', as in Wei et al. [2022] or Yao et al. [2023], that either enrich the context of the LLM, or provide disparate reasoning steps to compare, our approach can be seen as providing 'thoughts' for and against particular outputs, in the spirit of Miller [2023]."