MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning

2026-06-10 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

MoCA-Agent is a novel market-of-claims code agent designed to enhance accuracy in financial and numerical reasoning tasks, where precise grounding in facts, formulas, and units is critical. This system addresses the issue of subtle errors producing plausible but incorrect results by replacing free-form multi-agent debate with claim-level verification. MoCA-Agent decomposes questions into typed atomic claims, uses specialist "trader agents" to evaluate these claims, and synthesizes an executable Python program from the accepted evidence. A code-aware verifier then checks the program for consistency and errors, allowing for one market-aware repair round. Using a fixed Qwen3.6-27B backbone, MoCA-Agent achieved strong performance across ten public benchmarks, including 78.3% on FinQA, 76.0% on FinanceMath, 71.2% on MultiHiertt, 86.9% on ESGenius, and an 85.6% average on FinChart-Bench. This approach demonstrates that aggregating evidence at the atomic claim level significantly improves robustness in high-stakes numerical reasoning.

Key takeaway

For AI Scientists developing financial reasoning systems, you should consider claim-level verification to enhance accuracy and robustness. This approach, demonstrated by MoCA-Agent's strong benchmark performance, mitigates errors from misread cells or incorrect operations. Implement a system that decomposes questions into atomic claims and uses confidence-weighted evidence aggregation to synthesize verifiable code. This can significantly improve the reliability of your high-stakes numerical reasoning applications.

Key insights

MoCA-Agent uses a market-of-claims approach for robust financial and numerical reasoning.

Principles

• Decompose questions into typed atomic claims.
• Verify claims via confidence-weighted market decisions.
• Synthesize executable code from supported evidence.

Method

MoCA-Agent decomposes questions into typed atomic claims, uses specialist agents to "buy or sell" claims, clears orders into accept/reject decisions, synthesizes a Python program, and verifies it with a code-aware verifier.

In practice

• Apply claim-level verification to financial QA.
• Use multi-agent systems for evidence aggregation.
• Integrate code generation with verification loops.

Topics

• Financial Reasoning
• Numerical Reasoning
• Multi-Agent Systems
• Code Generation
• Claim Verification
• Qwen3.6-27B

Code references

• UBC-NLP/MoCA-Agent

Best for: AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Precision over Probability: Orchestrating AI Agents for High-Stakes Finance — Decoupling LLM reasoning from deterministic calculation is crucial for accuracy in high-stakes financial AI.
• I Directed AI Agents to Build a Tool That Stress-Tests Incentive Designs. Here’s What It Found. — AI agents can build tools to stress-test economic incentive designs, revealing hidden flaws and ambiguities.
• Discover and Prove: An Open-source Agentic Framework for Hard Mode Automated Theorem Proving in Lean 4 — Hard Mode ATP, requiring answer discovery before proof, reveals a significant gap between LLM reasoning and formal provers.
• Multi-Sourced, Multi-Agent Evidence Retrieval for Fact-Checking — WKGFC unifies structured KGs and web evidence via an LLM agent for adaptive, self-improving fact-checking.
• One Brain, Many Blind Spots — AI agents' confident hallucinations stem from training; multi-agent architectures provide essential verification for high-stakes applications.
• Verify Before You Commit: Towards Faithful Reasoning in LLM Agents via Self-Auditing — SAVeR improves LLM agent faithfulness by verifying internal beliefs before action, preventing propagation of unsound reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.