A finance AI passed its own checks while mixing up AMD and NVIDIA data
A new open-source prototype shows which parts of a finance assistant to trust with numbers, and why a checker that sees only the writer's evidence misses planted errors.
The short version
- Mistral-7B got 29% of 100 financial calculations right; Python got 100%.
- A live price quote cut price error from 210% to 0.00%.
- Hybrid retrieval reached 76% Hit@5 on 50 filing questions.
- In a case study, both answers passed validation yet put AMD sample data under NVIDIA's and Alphabet's names.
- A check against the original passage caught 77% of planted errors; a shared-evidence check caught 35%.
Why it matters
If you use AI for money questions, a source link and a passing check are not proof that a number is right. This paper shows that a checker looking only at what the writer saw can approve wrong answers. The safer habit is to compare the claim against the original document yourself.
A fluent answer is not a correct one
As arxiv.org reported, a research paper called FinVeriQA starts from a familiar complaint. Large language models (the technology behind chatbots) write smooth financial analyses, but they get ratios wrong, recall out-of-date prices, and send questions to the wrong tool while sounding sure of themselves.
The authors ask a practical question: in a finance assistant, which component should do which job? They answer it with an open-source prototype that works over SEC filings (the official reports US public companies must file) and live market data. It is a preprint, meaning it has been posted publicly but not yet peer reviewed.
Giving each job to the right part
The prototype is a pipeline, not one big model. A small decision model called Laya reads each question, picks the right route, and screens out requests for personal financial advice. Plain Python code does every calculation. Retrieval (searching documents) and tools fetch the evidence, and Mistral-7B, an open language model, writes an answer with citations.
A code validator then checks two things: that every cited ID really exists, and that the numbers in the answer match the ones Python computed. The whole pipeline is exposed through a Model Context Protocol (MCP) server. MCP is a standard way for AI systems to connect to outside tools and data.
What the live tests showed
The numbers make the case for splitting the work. Laya routed questions better than Mistral-7B, 0.85 versus 0.52, and 0.86 versus 0.68 overall. The authors report a p-value of 0.0003 for this, which means the gap is very unlikely to be chance. A larger 12B Mistral model decided no better. DeepSeek-R1-8B did better, 0.79 versus 0.68, but ran about 9.9 times more slowly.
On arithmetic, Mistral-7B got 29% of 100 financial calculations right, while Python got all of them. Fetching a live quote cut a price error from 210% to 0.00%. Hybrid retrieval, which mixes keyword and meaning-based search, reached 76% Hit@5 on 50 filing questions. Hit@5 means the right passage appeared among the top five results.
- Lesson one: let code do the math.
- Lesson two: fetch current data instead of trusting a model's memory.
When the checker was fooled
The most striking part is a failure. In a live case study, both agent answers passed the validator. Yet they described AMD sample data under the names of NVIDIA and Alphabet. Every citation existed and every number matched the notes, but the notes themselves were wrong.
To explain this, the authors ran a controlled test on a benchmark called FinRank. They corrupted one figure in the evidence that the agents shared. The agents' agreement with each other stayed between 0.55 and 0.59, so they looked confident. Accuracy, meanwhile, fell from 0.50 to 0.00. Agreement among copies of the same model proved nothing, because they were all reading the same bad evidence.
Verify against the original
The authors compared two kinds of check. One looks at the same evidence the writer saw, which is what their validator does. It caught 35% of the errors. The other compares the answer with the original passage and caught 77%. For DeepSeek-R1-8B, the figures were 68% versus 90%.
The conclusion is that verification has to be independent of the evidence the writer used. A check that shares the writer's inputs can repeat the writer's mistakes. This is a result from a prototype and a controlled benchmark, not a verdict on commercial finance products.
What to watch
- Will other teams test checks against original documents in finance tools, rather than checks that reuse the writer's evidence?
- Do the results hold outside a prototype and a controlled benchmark, once the work is peer reviewed?
Try it
When an AI gives you a financial figure, ask which document and passage it came from. Then open that original filing or page and confirm the number and the company name yourself.
Sources and links
- Reportingarxiv.org: the article we sourced
First reported by arxiv.org. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.