Mint-Agent: Introducing Finance-Native Agentic Foundation Models
The paper introduces Mint-Agent, a family of finance-native agentic foundation models built on a specialized data engine, interaction harness, and advanced training pipeline that achieves state-of-the-art reliability and executability in complex, auditable financial tasks.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of finance, a correct answer is rarely just a single number or a simple fact. It is the result of a careful chain of events: finding the right document, identifying the correct time period, comparing values from different sources, and performing calculations that must hold up under scrutiny. When a human analyst makes a mistake in this process, they can usually retrace their steps to find where the logic broke. But when an artificial intelligence system attempts the same task, it often produces a plausible-sounding answer that is built on a hidden error, a misread table, or a source that was relevant but not authoritative. Without a clear, traceable path from the original evidence to the final conclusion, the result cannot be trusted or fixed. This gap between a confident answer and a verifiable truth is the central challenge researchers are trying to solve as they teach machines to act as independent financial investigators.
A team of researchers has introduced a new approach to this problem with a system they call Mint-Agent. Rather than simply training a computer to memorize financial facts or to guess the right answer, they built a system designed to treat every conclusion as a claim that must be backed by recoverable evidence. The system operates on three main principles: it learns from real-world financial documents to understand the specific rules of the trade; it uses a specialized environment that keeps a permanent, auditable record of every search and calculation it performs; and it learns through a process that rewards not just the final answer, but the correctness of the entire journey taken to get there. The result is a family of artificial intelligence models that can conduct long, complex financial research while maintaining a clear trail of evidence that proves how they reached their conclusions.
The researchers constructed their system by first building a massive library of financial tasks derived from real corporate reports, market data, and accounting records. They separated these tasks into two types. The first type involves atomic skills, such as extracting a specific number from a table or performing a standard calculation, where the evidence is already present. The second type involves long-horizon execution, where the system must find the evidence itself, often searching through multiple documents over an extended period to piece together a complex answer. To handle the second type, they created a digital environment called MintHarness. This environment acts as a secure workspace where the AI can interact with tools, search for information, and perform calculations. Crucially, this environment keeps a permanent ledger of every action taken, every document viewed, and every number calculated. This ledger ensures that even if the AI forgets a detail, the record of how it found the answer remains intact and can be reviewed by a human.
To teach the AI how to use this environment effectively, the researchers employed a training method that separates the skills of financial reasoning from the skills of managing a long research project. They first trained one version of the model to be an expert at pure financial reasoning, focusing on getting the math and logic right when the data is already in front of it. They trained a second version to be an expert at execution, focusing on how to search, when to stop, and how to organize information over many steps. Finally, they merged these two experts into a single, unified system. This combined model, which comes in two sizes, was tested against a wide range of professional financial benchmarks. The larger model, named Mint-Ag, achieved a score of 98.33 on a test of financial reasoning, outperforming several leading commercial systems. On tests requiring complex, multi-step research, it also achieved the highest scores, reaching 76.00 on one major benchmark and 60.49 on a newer, more difficult version. The smaller model, Mint-Cu, which is much more compact, also performed exceptionally well, scoring 69.86 on a search-heavy task, proving that strong financial research capabilities can be retained even in a smaller system.
What makes these results significant is not just the high scores, but the nature of the performance. The researchers found that their models do not just produce answers; they produce answers that can be audited. In one detailed example, the system successfully tracked a forecast for a semiconductor company through thirty-seven steps of investigation. Along the way, it encountered blocked documents, incorrect time periods, and conflicting data. Because the system maintained a persistent record of its findings, it was able to discard the dead ends, correct its own mistakes, and recompute the final numbers using only the valid evidence it had secured. In another case involving a corporate acquisition, the system waited to lock in a final price only after finding the definitive legal document, ignoring earlier, ambiguous sources. This ability to distinguish between a lead and a fact, and to keep a clear record of the difference, is what allows the system to be trusted.
The researchers argue that for artificial intelligence to be useful in high-stakes fields like finance, it must move beyond simply generating text that looks correct. A trustworthy agent must be able to show its work, respect the time boundaries of the data it uses, and be willing to revise its conclusion if the evidence does not support it. By building a system where data construction, execution, and learning are all tied to a single standard of verifiable evidence, the Mint-Agent team has demonstrated that it is possible to create financial intelligence that is not only capable but also accountable. The findings suggest that the future of reliable AI in finance lies not in making models bigger or faster, but in engineering them to leave a clear, testable trail of how they reached every decision.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.