A check against the source text caught at most 77 per cent of Mistral-7B’s errors in a test
A check that reads the same material as the language model that wrote the answer lets through a large share of wrong figures, and even data for the wrong company, according to a preprint by researchers at San José State University.
On 18 November, researchers at San José State University published a preprint on arXiv about FinVeriQA. The system is an open-source research prototype that answers questions about US companies’ SEC filings and market data.
FinVeriQA splits the work. A small decision model, Laya, chooses which agents to run for a question and refuses requests for investment advice. Python calculates every figure, and the Mistral-7B language model writes a sourced answer. Finally, code checks that the cited sources exist and that the figures match the calculated ones.
The division of labour rests on measurements. Mistral-7B got 29 of 100 financial calculations right; Python got all of them. Without a live share price, the share prices the model gave differed from the real ones by 210 per cent on average; with it, by 0.00 per cent.
Passing the check still did not guarantee a correct answer. When the system was asked how Nvidia’s revenue and margins had changed, Laya chose the right route and every figure matched Python’s calculations. Yet the answer described the AMD sample data bundled with the system under Nvidia’s name. The checker compares the answer only with the notes it was written from. The case study had five questions, so it shows the types of error but not how common they are.
The researchers tested the mechanism on the FinRank dataset by replacing one figure in the source text with a plausible wrong value. Self-consistency, the agreement among five repeated answers, did not weaken (0.55 → 0.59), even though the share of correct answers fell from 0.50 to zero. The answers stayed just as consistent because the model read the same wrong figure each time. The reasoning model DeepSeek-R1-8B did not find the correct value either.
The table shows what share of the answers relying on the wrong figure each check flagged as incorrect.
| Model | Same material as the writer | Original source text |
|---|---|---|
| Mistral-7B (60 questions) | 35% | 77% |
| DeepSeek-R1-8B (40 questions) | 68% | 90% |
Source: FinVeriQA preprint, FinRank dataset, 4-bit models, single run.
The check against the original source text had access to error-free text, so its figures are an upper bound, not the results of a finished checker. The two models’ figures cannot be compared directly because the samples differ in size.
The authors argue that the check must be independent of the material the writer saw. Each figure should therefore be retrieved separately from the filing or its machine-readable XBRL data. They also plan to check that the notes concern the company asked about, that the source supports each sentence and that every part of the question is answered. These are only plans so far, and the prototype’s agents still partly use sample data.
The code and evaluation data are in the GitHub repository SJSU-CMPE259/AI-Stock-Research-and-Earnings-Assistant, the researchers said.