Does Grounding Argument Strengths in Evidence Text Help Claim Verification? A Study of Token-Anchored Arguments vs. ArgRAG

This paper investigates whether grounding argument strengths in evidence text improves automated claim verification. Building on ArgRAG, the study introduces the Token Argument Model (TAM), which represents retrieved evidence as shorter arguments anchored to verbatim text and estimates argument strength using a Natural Language Inference (NLI) model. TAM is evaluated across four experimental settings using PubHealth and SciFact. Results show that grounding argument strengths does not consistently improve claim-verification accuracy over uniform argument strengths, and randomly shuffled strength scores achieve comparable results. However, representing evidence as shorter, quote-anchored arguments improves accuracy by 6.0 percentage points on PubHealth in the top-5 retrieval setting. The findings suggest that improving argument representation can be more beneficial than assigning model-derived strength scores. Reproducibility note: The reproduced ArgRAG baseline does not match the accuracy reported in the original paper. Results should therefore be interpreted as relative comparisons within this experimental setup, not as direct comparisons with the original benchmark performance.