Research auditing step-level credit assignment in LLM agents finds that LLM-judge scores, logprob ratios, and confidence signals fail to identify which steps actually matter better than chance when...
New benchmark evaluates how well LLMs retrieve evidence across invoices, POs, and ledgers for financial root-cause analysis—separating true reasoning from document access.
LEDGER introduces claim-to-evidence trace graphs to make LLM agent outputs auditable, addressing the challenge of verifying correctness in autonomous workflows that execute code, edit files, and ge...
Researchers audit three self-evolving financial AI agents (SkillOpt, AWM, ReasoningBank) in e-banking scenarios, finding that performance gains don't guarantee security preservation or behavioral c...
Research prototype binds AI policy decisions to cryptographically signed, durable audit records with explicit trust boundaries—critical infrastructure for AI systems in regulated accounting/audit e...
Fiducia-bench, a new research benchmark, shows that decomposing financial agents into components undermines governance—reducing escalation, abstention, and auditability compliance critical for regu...
Research on ProcessBench traces reveals LLMs soften audit verification thresholds by 2.8–11.5 percentage points when prior repair episodes are in context, reducing false alarms across 15 model vari...
AgentRelBench, a new reliability framework, tested nine LLM models on accounting operations and found systematic failures in irreversible tasks—exposing that single-run audits miss critical agent d...
Researchers propose Mandato, a protocol-level governance framework enforcing digitally signed mandates on AI agent actions with cryptographically chained audit trails—addressing authorization and a...
CLAIR-Fin, a nine-agent framework, verifies financial claims at the atomic level using a typed ledger and cross-modal debate to prevent hallucinations in retrieval-augmented financial QA systems.
Research proposes Matrix, a causal-state layer for agentic workflows that enforces governed execution with inspectable provenance—critical for institutional settings like accounting where correct o...
Researchers propose policy-as-logic framework combining LLMs for fact extraction with formal logic solvers to ensure generative AI systems correctly apply tax rules and policies—making responses au...
Researchers develop GRPO-fine-tuned language model that outperforms GPT-4 and Claude on financial advice generation from business records, using reinforcement learning to improve numerical reasonin...
Researcher argues agentic AI governance in financial services hinges not on capability but verifiability—the gap between authority demands and explainability retained after automated decisions.
Research paper argues agentic AI governance in fintech is constrained by verifiability gaps—the shortfall between delegation authority demands and explainability/reproducibility—not capability limits.
Researchers propose 'Rule of Thumb' explainability framework to interpret AI decisions using partial information, with applications to LLM classification and AI system audits—relevant for accountan...
arXiv research proposes unified authentication gateway for Model Context Protocol servers connecting LLM agents to enterprise tools, addressing governance gaps across dozens of internal implementat...
Researchers propose SAGE-Fin, a structured authority-handoff contract system that prevents AI financial agents from executing unauthorized transactions by controlling proposed effects at runtime ra...
Researchers argue AI model evaluation for accounting must measure verification effort, not just correctness—identifying a new failure mode where auditors miss errors within budget constraints.
Research framework for auditing AI recommendation systems by comparing outputs against complete data census, addressing critical accountability gaps in AI-driven decision-making for high-stakes dom...