Research on 99,952 examples shows correct domain rubrics improve LLM evaluator accuracy by 2.11 points, suggesting specialization matters more in evaluation rules than model weights.
Researchers introduce FinanceHarness, an agentic AI framework designed specifically for financial deep research that combines LLMs with layered agent architecture and point-in-time benchmarks to au...
arXiv paper introduces ProofAgent Index (PAI), a governance readiness framework for AI agents in production, addressing gap between capability demos and real-world operational constraints in accoun...
Researchers propose Hierarchical Reranker, a RAG framework designed to improve how AI systems retrieve and analyze large-scale financial documents like 10-K filings and tabular disclosures, address...
Mercor and Ramp launch APEX-Accounting, a 160-task benchmark evaluating whether frontier AI models can perform core accounting work like reconciliation, accrual, transaction posting, and reporting.
Mercor and Ramp release APEX-Accounting, a research benchmark with 160 expert-authored tasks assessing whether frontier AI models can perform actual accounting work like reconciliations, accruals, ...
GAUGE benchmark evaluates agent-built financial forecasts using multiple independent analyst models rather than single reference, addressing the challenge that reasonable financial projections admi...
Researcher audits 40 empirical records on generative AI in accounting, finding that model obsolescence and publication lag create systematic validity gaps—proposing a claim-currency framework to as...
Academic research using real estate case studies to help firms distinguish between promising and doomed AI projects, with frameworks for ROI estimation and project selection.
New authorization audit method evaluates how LLM agents select tools and arguments, addressing risks when agents rely on untrusted or unauthorized evidence in accounting and compliance tasks.
Academic paper examines how declared vs. implemented identity rules in record systems can diverge without triggering audit detections, with implications for financial data integrity and audit proce...
New evaluation framework tests AI systems on multilingual financial question answering across English, Chinese, Japanese, Spanish, and Greek with 256 benchmark items.
Researchers combine gradient boosting, graph features, SHAP explanations, and bounded LLM agents to build auditable fraud detection that scales with transaction volume while remaining reviewable fo...
Researchers propose robust financial statement fraud detection framework using LLMs that addresses overfitting in prior work by testing generalization across companies and time periods, advancing f...
Researcher proposes verifiable event timeline for autonomous commerce agents with tamper-evident auditability and fraud detection, addressing gaps in AP2 and ACP protocols.
Researchers propose using misinformation detection and explanation techniques to assist financial auditors in identifying material misstatements in balance sheets, income statements, and cash flow ...
Researchers propose DRQN-CMDP, a deep reinforcement learning approach using recurrent neural networks to adaptively schedule audits of off-chain data while maintaining security miss-rate constraint...
Leni production agent achieves higher reliability through verification loops with lightweight task-specialized models, evaluated across spreadsheet and accounting benchmarks.
Systima's testing reveals Claude's subagent architecture wastes 2–6x input tokens and offers no speed gains over sequential task processing, raising cost concerns for AI automation workflows.
Kimi K3 evaluated on TaxCalcBench for ability to file American tax returns, testing cross-border AI capability in specialized tax domain.