Presents ArgLLM-App, a web system using argumentative LLMs for explainable, contestable decisions with human interaction.
ArgLLM-App is a web-based interactive system that implements Argumentative Large Language Models (ArgLLMs) to support explainable and contestable decision-making through computational argumentation. It enables users to engage with binary decision tasks, such as claim verification, by visualizing the reasoning process in the form of quantitative bipolar argumentation frameworks (QBAFs). These QBAFs are represented graphically, with arguments as nodes and attack or support relations as edges, allowing users to inspect, modify, and contest the model's reasoning.
The system allows customization of QBAF depth (1 or 2), breadth (up to 4 attackers and supporters per argument), and gradual semantics such as DF-QuAD, Euler-based, and Quadratic Energy for evaluating argument strength. To avoid cognitive overload, depth is limited to 2. Users can adjust base confidence scores via sliders, add new attackers or supporters to arguments, or upload trusted external sources in PDF format to inform the reasoning process, aligning with Retrieval-Augmented Generation (RAG) principles.
Interaction can occur either through direct manipulation of the QBAF visualization or via a chat interface, enabling human-in-the-loop refinement of AI-generated arguments. ArgLLM-App is publicly accessible at argllm.app and includes a demonstration video at youtu.be/vzwlGOr0sPM. The system currently uses OpenAI's LLMs, though future extensions envision support for multi-agent setups with diverse LLMs and concurrent multi-user collaboration.
ArgLLM-App advances human-AI collaboration by making AI decisions not only interpretable but also modifiable and contestable, enhancing trust and transparency in AI-assisted decision-making. It will be presented at the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026), taking place from May 25 to 29, 2026, in Paphos, Cyprus
ArgLLM-App introduces a web-based interactive system designed to integrate Large Language Models (LLMs) with structured argumentative reasoning to enhance transparency in decision-making. Unlike standard LLM interfaces that often generate opaque conclusions, ArgLLM-App compels the underlying model to produce outputs mapped to formal argumentation frameworks, decomposing reasoning into explicit claims, premises, and warrants. This structured approach transforms the LLM from a black-box oracle into an agent that can expose its logical chain, allowing users to trace the evidence and rules used to arrive at a specific decision.
The system’s primary contribution lies in its emphasis on "contestability" and human-in-the-loop collaboration. Rather than simply accepting an AI's output, users can interact with the argument structure, challenging specific premises or requesting alternative conclusions based on modified assumptions. This interaction model facilitates a dialectical process where the human user can scrutinize, refine, and correct the AI's reasoning in real-time. By utilizing argumentation schemes to guide the LLM's generation, the system ensures that the outputs are not only coherent but also adhere to logical standards that are easier for humans to validate.
This research is significant for advancing Explainable AI (XAI), particularly in high-stakes domains where accountability is paramount. As LLMs move toward supporting critical decisions in fields like law, healthcare, or policy, the inability to audit or challenge their reasoning remains a major barrier to adoption. ArgLLM-App demonstrates a practical architecture for aligning LLM capabilities with human cognitive needs, proving that argumentative reasoning can be operationalized effectively within generative AI systems to foster trust and collaborative intelligence.
The paper introduces ArgLLM-App, a web-based system designed to enhance human-AI collaboration in decision-making by leveraging large language models (LLMs) for explainable, contestable argumentation. Unlike traditional LLM applications that provide opaque or one-sided outputs, ArgLLM-App facilitates interactive reasoning by generating structured arguments with premises, claims, and counterarguments. Users can challenge, refine, or edit these arguments in real time, enabling a dynamic dialogue where the AI adapts its reasoning based on human feedback. The system is built on argumentative LLMs, which are fine-tuned to produce coherent, logically connected justifications rather than surface-level responses. This approach addresses a key limitation in AI decision-support tools—lack of transparency and contestability—by embedding reasoning within a framework that users can scrutinize and modify.
The paper’s key contributions include: 1. A novel interactive framework for human-AI argumentation, where users engage in a back-and-forth to refine AI-generated reasoning. 2. Explainability through structured argumentation, moving beyond post-hoc explanations to integrate reasoning into the decision-making process itself. 3. Contestability mechanisms, allowing users to dispute, edit, or request alternative arguments, which helps identify biases or weaknesses in AI outputs. 4. Technical implementation details, including the use of prompt engineering and fine-tuning to ensure LLMs generate high-quality, contestable arguments.
Why this matters: As AI systems increasingly aid in high-stakes decisions (e.g., legal, medical, or policy domains), the need for interpretable, collaborative reasoning grows. ArgLLM-App demonstrates a practical step toward human-centered AI, where transparency and user agency are prioritized. By framing AI assistance as an argumentative dialogue rather than a black-box oracle, the system could improve trust, accountability, and the overall quality of AI-driven decisions. This work is particularly relevant for researchers in human-AI interaction, explainable AI, and argumentation mining, as well as practitioners developing decision-support tools where reasoning must be both auditable and adaptable.
Source: [ArgLLM-App on arXiv](https://arxiv.org/abs/2602.24172)