EvidenceLens: A Visual Tool for Financial Question Auditing

In the realm of finance, large language models are increasingly utilized to interpret complex documents such as annual reports and earnings presentations. However, the challenge lies in verifying the accuracy of their responses, especially in critical financial contexts. Often, the answers generated can mix well-supported facts with vague inferences and unverified assertions, making it hard for analysts to trust the information. EvidenceLens addresses this issue by providing a visual analytics tool that redefines financial question answering as a process of aligning claims with their corresponding evidence. This innovative system breaks down responses into individual claims, evaluates the strength of their support, identifies gaps in evidence, and allows for detailed examination of each claim alongside relevant text, tables, and charts. The centerpiece of EvidenceLens is a claim-evidence matrix that highlights areas of coverage, contradictions, and inconsistencies. To enhance reproducibility, the system includes a JSON-based schema and a structured alignment pipeline, ensuring that the auditing process is both transparent and systematic. Through practical auditing scenarios, EvidenceLens empowers financial analysts to differentiate between substantiated claims and overconfident assertions that traditional chat interfaces often obscure.