arXiv paper introduces protocol for constructing verifiable LLM agent tasks that isolate domain-specific knowledge from task execution, enabling better evaluation of whether agents truly understand...
Researchers introduce Oculi, an agentic conversational platform that automates credit risk analysis by translating natural language queries into SQL, statistical tests, and visualizations—reducing ...
New 8B-parameter LLM VERA-8B identifies audit risks in SEC filings with evidence-grounded reasoning, addressing the gap between general financial language models and audit-specific requirements for...
Researchers develop validated synthetic data pipeline and semantic-aware fine-tuning method to improve LLM performance on financial question answering tasks involving tables, charts, and narrative ...
Researchers propose architecture pattern separating LLM agent personas from execution to maintain audit trails while allowing instruction evolution—key for deploying AI agents in governed organizat...
Pre-registered research reveals that the dominant statistical method for auditing LLM judges—difference-in-differences on rating scales—produces spurious findings due to censoring effects that conf...
Academic paper introduces Trace Integrity criterion for LLM data agents in structured-data tasks, enabling auditable reasoning paths critical for accounting compliance and financial systems.
Researchers introduce findr, a machine learning framework that combines predictive accuracy with interpretable credit risk decisions and fairness constraints for regulatory compliance.
FraudBench introduces protocol-sensitive evaluation framework for assessing adversarial robustness of ML fraud-detection systems in financial risk assessment, accounting for domain constraints and ...
Researchers compare deterministic capital gains engines vs. RAG-augmented systems in multi-agent financial advisory, finding structured computation outperforms retrieval for personalized tax optimi...
arXiv research proposes POLANYI++, a logic-augmented generation method to prevent LLMs from reinforcing fabricated premises in accounting outputs, addressing hallucination risks in financial report...
Researchers introduce FinixDoc, an end-to-end agentic parsing system with a 4B vision-language model, to address gaps between benchmark and real-world financial document parsing accuracy and struct...
Academic paper introduces HANSARD, a reference architecture for auditing autonomous AI agents in finance and security, addressing liability gaps when multi-agent systems cause harm through provenan...
Researcher demonstrates single person directing fleet of AI agents via machine verification at scale, flipping traditional cost-overhead model where verification becomes essential productivity safe...
Researchers present SENTRY, a deterministic ML pipeline that automates risk assessment for IT changes in financial institutions, replacing error-prone self-reported questionnaires with auditable, c...
Academic paper introduces methodology to assess whether AI agents reach correct accounting judgments through valid reasoning—critical for audit, compliance routing, and training feedback systems.
BC-Bench evaluates agentic AI systems on real-world Microsoft Dynamics 365 Business Central tasks, revealing gaps between general AI performance and domain-specific ERP accounting work.
ReguSim benchmark reveals DeepSeek and Gemini agents cite compliance rules but still violate constraints in trading scenarios, highlighting gaps between LLM reasoning and execution.
NVIDIA's Nemotron 3.5 Lightning LLM ranks near bottom of DualEntry's accounting benchmark despite speed and cost advantages, scoring just 45.5% on accounting-specific tasks.
AICPA releases revised digital assets practice aid covering stablecoins, mining revenue, and current auditing standards for evolving crypto accounting landscape.