[论文解读] ArgLLM-App: An Interactive System for Argumentative Reasoning with Large Language Models

Adam Dejl, Deniz Gorur+1 more|arXiv (Cornell University)|Feb 27, 2026

Explainable Artificial Intelligence (XAI)被引用 0

一句话总结

tldr: ArgLLM-App 是一个基于网页的系统，实现 ArgLLMs 用于二元决策，提供交互式 QBAF 可视化、用户编辑和基于文档的 RAG 指导来解释和辩论推理。

ABSTRACT

Argumentative LLMs (ArgLLMs) are an existing approach leveraging Large Language Models (LLMs) and computational argumentation for decision-making, with the aim of making the resulting decisions faithfully explainable to and contestable by humans. Here we propose a web-based system implementing ArgLLM-empowered agents for binary tasks. ArgLLM-App supports visualisation of the produced explanations and interaction with human users, allowing them to identify and contest any mistakes in the system's reasoning. It is highly modular and enables drawing information from trusted external sources. ArgLLM-App is publicly available at https://argllm.app, with a video demonstration at https://youtu.be/vzwlGOr0sPM.

研究动机与目标

• Motivate the use of ArgLLMs to achieve explainable and contestable decision-making.
• Provide a modular web system that generates QBAFs from LLM-derived base scores for binary decisions.
• Enable user interaction to adjust base confidences and expand the QBAF through attackers/supporters.
• Support document-based and chat-driven augmentation of QBAFs to incorporate trusted sources.

提出的方法

• Implement ArgLLM logic on a server mediating access to a base LLM.
• Represent decisions as QBAFs with attacks and supports between arguments.
• Compute final claim confidence via gradual semantics (e.g., DF-QuAD) with configurable depth and breadth.
• Allow user-controlled adjustments of base scores and addition of attackers/supporters.
• Incorporate document-based QBAF generation by parsing PDFs into prompt data for the LLM.

实验结果

研究问题

• RQ1 How can ArgLLMs produce transparent, contestable decisions for binary tasks?
• RQ2 What is the effect of depth, breadth, and chosen gradual semantics on QBAF-derived conclusions?
• RQ3 Can users effectively revise outputs by editing base confidences and adding new arguments or sources?
• RQ4 What is the feasibility and benefit of integrating external documents via RAG-based QBAF generation?

主要发现

• ArgLLM-App supports depth 1 or 2 and breadth up to 4 attackers and 4 supporters per argument.
• Users can modify base confidences via a slider to reflect trusted evidence strength.
• Users can augment the QBAF by adding supporters or attackers through interactive controls.
• The system allows document-based QBAF generation by parsing uploaded PDFs and integrating them into LLM prompts.