An LLM-Based System for Argument Reconstruction

The paper doesn't just extract arguments; it maps their logical relationships (support, attack, undercut) as a structured graph. That relational layer is what makes the output actionable for downstream tasks rather than merely descriptive.

This work sits at the intersection of two recent Modelwire themes. The 'Negation Neglect' paper from this week showed LLMs struggle to internalize negated claims during training, suggesting they may conflate mention frequency with truth. Argument reconstruction directly addresses that gap by making negation and logical opposition explicit as graph structure rather than relying on implicit token-level understanding. Separately, the WARDEN transcription work demonstrated that decomposing end-to-end tasks into interpretable substeps improves performance in low-data regimes. This argument pipeline follows the same principle: breaking natural language into premises, conclusions, and their logical bonds creates intermediate representations that should be more learnable and auditable than black-box end-to-end reasoning.

If this system correctly reconstructs arguments containing explicit negations or contradictions at >85% accuracy on a held-out test set where the baseline LLM (without graph structure) scores <70%, that would validate whether structured decomposition actually solves the negation internalization problem. If it doesn't, the bottleneck lies elsewhere in the reasoning pipeline.