MOCA-Agent: Redefining Trust in Financial AI with Claim-Level Precision

In the area of financial and tabular question answering, precision is key. A single error in reading data or executing operations can lead to wildly incorrect results that appear plausible at first glance. Enter MOCA-Agent, a novel approach designed to ensure accuracy by focusing on claim-level verification rather than broad, multi-agent debates.

Claim-Level Verification: A Game Changer?

MOCA-Agent innovates by breaking down complex questions into atomic claims. These claims are then subjected to a market of specialist trader agents who weigh in with buy or sell decisions, transforming these into confidence-weighted outcomes. The system synthesizes a Python program backed by this market-supported evidence, checked meticulously by a code-aware verifier for consistency and typical financial reasoning errors. Importantly, this process allows for only one market-aware repair round, aiming to maintain stringent accuracy.

But why is this important? Traditional systems often rely on broader debate mechanisms that can miss the mark exact financial reasoning. By focusing on claim-level verification, MOCA-Agent enhances robustness and reliability in numerical reasoning, a critical need in high-stakes financial environments.

Impressive Benchmark Results

The benchmark results speak for themselves. MOCA-Agent, using a Qwen3.6-27B backbone, achieves remarkable performance across ten public benchmarks. Notably, it scores 78.3% on FinQA, 76.0% on FinanceMath, and an impressive 86.9% on ESGenius. Compare these numbers side by side with traditional methods, and the advantage is clear. Aggregate evidence at the atomic claim level offers a more accurate and reliable foundation.

What the English-language press missed: the significance of these results goes beyond mere numbers. It sets a new standard for AI in financial reasoning, where precision and reliability aren't just desired but essential.

Looking Ahead

MOCA-Agent's approach begs the question: are we witnessing the future of AI in financial question answering? As systems become more complex, the need for accuracy becomes ever more critical. MOCA-Agent represents a shift towards more granular, reliable AI, ensuring that results aren't just plausible but verifiably correct. In doing so, it tackles one of the biggest challenges in financial AI today.

The paper, published in Japanese, reveals a promising direction for AI development. By refining the focus from broad answers to specific claims, MOCA-Agent not only enhances accuracy but also builds trust in AI-driven financial analysis. This shift could be the key to unlocking new levels of performance and reliability in financial and tabular reasoning tasks.