Academic paper proposes anytime-valid statistical guarantees for adaptive auditing of generative AI systems, addressing the bottleneck of costly manual annotation in characterizing AI failure modes.
Researchers introduce FakeWiki benchmark and "pinpoint provenance" method to help auditors identify which source documents support LLM responses, addressing a key challenge in AI accountability for...
Researchers release AuditRepairBench, a 576K-cell corpus revealing how evaluator configurations cause AI repair leaderboards to reorder, addressing methodological failures in agent-based accounting...
Researchers present automated framework using Snowflake Document AI to extract structured data from millions of unstructured PDF statements, replacing labor-intensive manual audit sampling with sca...
Research identifies a critical vulnerability in AI auditing: systems verbally agree to process instructions but systematically disobey them, creating risks for financial audits relying on procedura...
Research reveals OpenClaw, a widely-used agentic-AI gateway, catches zero of four critical audit failures (gate-bypass, audit-forgery, silent host failure, wrong-target) on a 1,600-sample test—expo...
Semia framework uses constraint-guided synthesis to audit LLM agent skills—hybrid code-prose configurations—addressing blind spots in conventional static analyzers when agents execute financial or ...
Research introduces CA-ThinkFlow, an LLM framework designed to handle jurisdiction-specific chartered accountancy tasks including multi-step numerical reasoning and regulatory compliance without re...
Claw-Eval-Live introduces a live benchmark for testing LLM agents on evolving accounting and business workflows, addressing limitations of static task sets in existing evaluations.
Claw-Eval-Live introduces a live benchmark for evaluating LLM agents performing real-world accounting and business tasks across multiple software tools, addressing limitations of static benchmarks.
Academic research exposes how rubric wording and metric choice skew NLP benchmarks used to evaluate LLMs on earnings calls and financial disclosures, risking poor model selection for financial repo...
Researchers propose a trusted execution environment approach to make LLM-based grant evaluation auditable and contestable while protecting model and scoring logic from applicant gaming.
Research benchmarks Claude Opus, Sonnet, and GPT-5.4 on 100 accounting queries, showing semantic context layers reduce hallucination rates and improve answer accuracy for LLM-powered data analytics.
Researchers model audit policy design as a game theory problem to detect fraud in resource allocation, offering theoretical framework for auditors balancing fraud detection costs against misreporti...
New benchmark evaluates LLM reliability for deterministic tasks like invoice extraction and transcript parsing—critical for automation workflows in accounting and bookkeeping.
Researchers propose cryptographic audit trail framework to create verifiable, tamper-proof logs of AI agent actions, addressing compliance and accountability needs in regulated environments like ac...
Survey finds 74% of small business owners use AI tools monthly with 58% reporting increased happiness; CPAs seeing adoption of AI drive autonomy and resilience in practice management.
Deloitte survey finds cost management is CFOs' top internal risk, with automation and tech upgrades identified as the most effective mitigation approach.
arXiv paper addresses transparency gap in autonomous AI agents by developing methods to audit and control agent actions in spreadsheet environments, enabling users to oversee multi-step AI decision...
Researchers release IndiaFinBench, a 406-question evaluation dataset for testing LLM performance on Indian financial regulatory compliance and filings—addressing gap in non-Western financial NLP be...