EvidenceLens: A Claim-Evidence Matrix for Auditing Financial Question Answering

A new visual analytics prototype called EvidenceLens aims to help financial analysts verify answers produced by large language models by decomposing those answers into atomic claims and mapping them against source evidence, according to a paper posted to arXiv on June 19, 2026. [1][2] The system treats financial question answering as a claim-evidence alignment problem. It breaks an LLM-generated answer into individual claims, then summarizes the support composition, confidence levels, and support gaps for each claim. The core interface is a multimodal claim-evidence matrix that makes coverage, contradiction, and modality imbalance immediately visible. [1][2] Analysts can coordinate claim-level inspection with source passages, table cells, and chart regions. The authors argue that fluent LLM answers often blend directly grounded statements, weak synthesis, and unsupported claims across narrative text, tables, and charts, making verification difficult in high-stakes financial workflows. [2] The paper specifies a JSON-based artifact schema, a lightweight multimodal alignment pipeline, and a deterministic review-priority ranking that maps backend signals into an auditable visual structure. Through representative report-auditing scenarios, the researchers show how EvidenceLens helps analysts distinguish grounded claims from overconfident synthesis that conventional chat interfaces flatten. [2] The work appears on arXiv, an open-access repository of electronic preprints that is not peer-reviewed. As of November 2024, the repository was receiving about 24,000 submissions per month. [9] The paper is hosted within arXiv's computer science and information retrieval category. [1] EvidenceLens is listed as an arXivLabs project, a framework that allows community collaborators to develop and share experimental features directly on the arXiv website. [1][7] arXivLabs was formalized in 2020 to enable collaborations with individuals and organizations that share arXiv's values of openness, community, excellence, and user data privacy. [7] The framework appears as a set of tabs at the bottom of article abstract pages, highlighting tools such as the Bibliographic Explorer and the CORE Recommender. [7][8] arXiv has stated that third-party collaborators receive only minimal and anonymized data about users, strictly for ensuring correct feature functionality. [7] Large language models, the type of model EvidenceLens is designed to audit, are machine learning models with many parameters trained on vast amounts of text for natural language processing tasks such as language generation. [11] The EvidenceLens paper addresses a specific gap in applying such models to financial documents, where verification of outputs is critical. [2]