Neurosymbolic Learning for Inference-Time Argumentation

This paper introduces Inference-Time Argumentation (ITA), a new framework designed to improve how AI models verify claims in high-stakes fields like finance and healthcare. When information is incomplete or conflicting, a simple "true" or "false" answer is often insufficient. ITA addresses this by treating claim verification as a three-valued problem—True, False, or Uncertain—and using formal argumentation to provide a transparent, reliable explanation for every verdict. Instead of relying on "black-box" reasoning, ITA builds an explicit map of supporting and attacking arguments, ensuring the final decision is mathematically tied to the evidence presented.

How ITA Works

ITA functions as a neurosymbolic system, meaning it combines the flexible language generation of Large Language Models (LLMs) with the rigid, logical structure of formal argumentation. The framework consists of two main parts: an argument generator that identifies reasons for and against a claim, and a base score model that assigns an "intrinsic strength" to each of those reasons. By applying a mathematical technique called argumentation semantics, the system aggregates these scores to calculate a final strength for the claim. This strength is then converted into a verdict, with the "Uncertain" label reserved for cases where the evidence is too balanced to support or reject the claim.

Training Through Argumentation

Unlike previous models that use off-the-shelf LLMs to guess arguments, ITA is specifically trained to optimize its argumentative output. The researchers use two primary training strategies. First, they use a "weakly supervised" approach where the model learns to assign base scores to arguments by looking at the final, correct verdict. Second, they use reinforcement learning to train the argument generator. By providing feedback based on whether the generated arguments lead to the correct final classification, the system learns to produce more relevant and helpful evidence. This turns the argumentation process itself into a training signal, rather than just a final step.

Results and Reliability

The researchers tested ITA on three adapted datasets and found that it outperforms existing argumentative models while remaining competitive with standard, non-argumentative prediction models. The primary advantage of ITA is its "faithfulness." Because the final verdict is computed deterministically from an explicit structure of arguments and scores, the system provides a clear, inspectable trail of evidence. This avoids the common issue in modern AI where models provide "post-hoc" explanations—justifications that are generated after the fact and may not actually reflect how the model reached its decision. With ITA, the explanation is the foundation of the decision, not an afterthought.