Trust, but Verify: A Faithfulness Framework for Financial LLMs
TL;DR: Financial AI can give you the right number for the wrong reasons. This talk introduces an evaluation framework — currently in active development and research, co-created by one of our own teachers, Colby — that checks how RAG systems reach their answers: whether every figure traces back to the source, the math holds up, and the evidence is cited correctly. For anyone working on LLMs, RAG, evaluation, or finance.
Financial QA systems built on Retrieval-Augmented Generation (RAG) can pull from 10-Ks and annual reports to answer questions — but how do we know the model's numbers are actually grounded in the source documents, and not just plausible-sounding guesses?
Most evaluation today checks whether an answer matches a ground-truth label. That tells you if the final number looks right, but not whether the reasoning behind it holds up. In finance, a superficially correct answer built on an unsupported calculation can still lead to a materially wrong conclusion.
In this talk, Colby presents a new evaluation framework that goes beyond answer-matching to assess how financial AI systems reach their answers, across three dimensions:
Numerical Grounding — can every figure in the answer be traced back to the retrieved evidence?
Computational Faithfulness — are the arithmetic steps logically consistent with the source data?
Evidence Attribution — does the answer correctly cite the passages it relied on?
Come for a deeper look at faithfulness, provenance, and trust in financial RAG systems — and why getting the right answer isn't the same as getting it for the right reasons.
🎤 About the Speaker — Colby Wang
Colby (Ziyu) Wang is a Master of Science student at Toronto Metropolitan University, researching under Dr. Vivian Hu on applied Natural Language Processing and Reinforcement Learning. A University of Waterloo CS alum, his interests span not only building AI systems but also teaching AI topics to others. More about him: https://c12wang.wixsite.com/colbyziyuwangblog
🌐 About AI Scholars
A peer-driven, high-quality learning community for engineers, students, researchers, and AI practitioners.