Researchers test whether detailed procedural traces in LLM audit loops make judges gullible; analysis of 4,551 compliance judgments shows error detection stays near ceiling despite trace variations.
Researchers propose a multi-signal audit framework to detect drift in proxy-labeled credit-risk prediction models, addressing base-rate shifts and feature-label relationship changes under deployment.
Audit of frontier LLMs reveals shallow document-reading ability: models score well on standard benchmarks but accuracy drops sharply when evidence is buried, with increased hallucinations, tool cal...
Researchers evaluate open-source large language models with retrieval-augmented generation for automating ESG KPI extraction from 498 real-world EU corporate reports, addressing a gap in domain-spe...
Academic paper argues that counting equivalent pairs overstates audit coverage in transformation systems; proposes four metrics (edge count, effective contrast rank, population support rank, spectr...
Researchers introduce Crypto Accounting Bench, evaluating frontier and open-weight LLMs on reconstructing journal entries for crypto transactions across 118 tasks from 7 organizations.
Research testing Google Gemini 3.0 Pro on 450 planted document errors reveals LLMs suffer batch-size degradation and confident hallucination—critical vulnerabilities for automated audit applications.
New scattering layer called Deposon makes LLM multi-step reasoning auditable by binding each reasoning node to a conserved state, enabling machine-checkable verification of discarded reasoning path...
OntoKG-EQ addresses the audit and reproducibility crisis in equity analysis by replacing ad-hoc spreadsheets with a provenance-grounded knowledge graph that links analyst queries to evidence, times...
Researchers propose event-to-action process mining framework enabling AI agents to evaluate proposed actions against organizational controls, risk, and authority—extending traditional process minin...
Researchers propose VST, a typed messaging protocol for AI agents that creates verifiable, replayed audit trails in financial alpha discovery workflows, addressing traceability gaps in LLM-based mu...
Researchers propose SerenAI, an LLM-based state-transition system that outputs auditable, verifiable predictions instead of unconstrained text—addressing the need for traceable AI in legal, financi...
Probity benchmark reveals LLMs give inconsistent answers on identical accounting queries 60+ times; researchers identify systematic defect where evidence is hidden from model's visible context window.
Controlled research identifies systematic numerical hallucination in domain-adapted LLMs used for financial reporting, challenging assumptions about fine-tuning effectiveness for critical accountin...
WAND, a new unsupervised anomaly detector, identifies suspicious financial data points and explains which features triggered the flag—eliminating post-hoc explanation overhead for accounting fraud ...
Researchers release FinRAG-QA, a 999-question benchmark dataset for AI systems to answer questions about banks' financial statements, addressing the complexity of comparative analysis across jurisd...
Two-thirds of employees believe their firms use AI for efficiency gains rather than value creation, raising concerns about burnout and workload in accounting roles.
Developer shares methodology for building self-improving AI agents that convert natural language to database queries using semantic benchmarks, with potential applications in accounting automation.
Academic paper introduces ensemble margin and local prediction variability metrics to audit decision systems affected by predictive multiplicity (Rashomon effect), enabling better human review rout...
Researchers propose AnalysisBank, an AI pattern library that generates financial reports by retrieving expert analytical moves matched to data signals, shifting automation from structural templates...