Tauric Research
Perspectives

Evidence-grounded reasoning

  • Intelligence

Retrieval helps models read financial documents; expert judgment is still hard to teach.

Most of what moves a market is written down: filings, earnings calls, central bank statements, news. Language models can read all of it, and they write fluent answers about it. The question for financial research is how to make a model’s conclusions rest on evidence analysts can verify.

State of the field

Retrieval-augmented generation was introduced partly because models could not show the provenance of what they said (Lewis et al., 2020 (external site)). Since then, models have learned to act as well as read: ReAct (external site) (Yao et al., 2022) interleaves reasoning with calls to external sources, so that a conclusion can rest on what was looked up rather than on what the model remembers. In finance, where a wrong decimal is a wrong answer, figures should come from data and tools, not from recall.

Reading is the easier half. Bridgewater’s AIA Labs, working with Thinking Machines, found that frontier models reached the mid-70s in accuracy on expert financial tasks with their best prompts, short of the 80% that investors expect from a system they would trust. In their words, “an explicit prompt can only convey the intuition an expert is able to put into words, while the judgments that matter most are often the hardest to articulate” (Su et al., 2026 (external site)).

Principles

Every conclusion should come with its evidence and with an account of what is missing. The Bank for International Settlements counts the tendency of AI models to hallucinate among the factors that amplify their risks (BIS, 2024 (external site)), so a claim that cannot be tied to a source should be softened or flagged. We would rather a model say it does not know than sound sure.

Our approach

Our agents read the same point-in-time data the Tauric API serves. We aim for models that compute figures with tools, tie each claim to its source, and keep a visible line between evidence and judgment.

References