Researchers develop LoRA-MINT to audit training data in fine-tuned LLMs
New methodology LoRA-MINT enables auditing of training data membership in domain-adapted LLMs, addressing IP and data sensitivity concerns for accounting applications.
The Wire — Page 3
17 more from arXiv: AI + Accounting
Research team proposes evidence tracing for LLM agent auditing
Academic paper introduces methods to trace and verify LLM agent decision-making through execution provenance, addressing audit and compliance needs for autonomous AI systems in high-stakes domains.
Provably Auditable LLM Agents Use Typed Lambda Calculus for Guaranteed Correc...
Agentic Redux, a new LLM agent architecture, uses typed lambda calculus to guarantee semantic correctness with append-only ledger auditability, demonstrated in healthcare billing compliance and sec...
New benchmark tests AI agents on auditable financial research workflows
Researchers introduce BigFinanceBench, a 928-item benchmark designed to evaluate financial-research AI agents on auditability and workflow transparency—measuring not just final answers but the deri...
Knowledge-Graph Pattern Brings Persistent Audit Trails to Compliance Detection
Academic paper proposes Violation Situation Pattern (VSP), a knowledge-graph approach that converts transient compliance violations into persistent audit objects with lifecycle tracking, review sta...
Study exposes blind spots in LLM benchmark contamination detection
Academic research identifies distribution shift and scale as failure modes in contamination detection methods for auditing LLM training data, raising questions about the reliability of current asse...
AuditFlow: Multi-Agent Framework Automates Structured Financial Audit Verific...
Researchers propose AuditFlow, a graph-grounded multi-agent AI system that automates financial audit verification by linking reported facts to taxonomy concepts and recomputing values against audit...
Financial LLM Agents Need Better Memory to Handle Market Complexity
Researchers propose an interaction-native knowledge system for financial AI agents that retain user context, preferences, and market assumptions to reduce errors and improve auditability in trading...
Unified guardrail system automates compliance checks for financial dispute do...
Researchers propose a guardrail orchestration layer that consolidates PII redaction, content moderation, and format validation for high-stakes financial documents like audit summaries and dispute n...
Framework optimizes LLM guardrails for evidence documents in dispute workflows
Production study demonstrates HOPM framework for auditable, evidence-grounded document generation in marketplace disputes—bridging AI reliability with accounting/compliance audit trails.
LLMs struggle with financial math; researchers propose data-centric fix for Q...
Researchers address numerical hallucinations in LLM-powered financial Q&A systems via data-centric compilation, improving reliability for high-stakes accounting applications where accuracy is criti...
LLM agents in accounting face hidden prompt injection risks in tool descriptions
Academic research reveals tool-augmented LLM agents used in accounting contexts are vulnerable to prompt injection attacks through multiple surfaces beyond tool outputs, challenging current securit...
Researchers expose LLM token billing auditing gaps as providers hide pricing ...
Academic research reveals that per-token LLM billing lacks auditability safeguards, enabling dishonest providers to overcharge—a critical issue for accounting firms using AI tools at scale.
New benchmark tests LLMs on financial statement verification accuracy
FinVerBench evaluates 15 LLMs on numerical consistency checks in SEC 10-K filings, introducing error taxonomy across arithmetic, linkage, and magnitude categories to measure AI readiness for financ...
Agentic AI pipeline reduces hallucinations via semantic caching and nested le...
Research on HOPE-inspired nested learning with semantic caching cuts LLM hallucinations in multi-agent pipelines, critical for reliable accounting automation systems.
Framework enables auditable AI decision-making with uncertainty routing for c...
Research paper introduces EvaluatorDPT, a system for production AI to handle uncertain cases through policy-governed abstention and real-time steering, addressing audit requirements in high-stakes ...
Researchers propose framework for measuring agentic AI technical debt in acco...
Academic paper introduces formal model distinguishing agentic technical debt from stochastic tax, offering accounting teams metrics to measure and simulate AI agent governance risks and operating c...
Researchers propose neuro-symbolic verification to catch LLM hallucinations i...
Academic paper presents hybrid formal logic + neural method to verify LLM outputs in high-stakes domains like accounting, addressing hallucinations and consistency risks.
AI adoption among tax firms nearly doubles in one year, survey finds
Blue J and CPA.com survey shows tax practitioners moving beyond experimentation to embed AI as core workflow tool, signaling shift from pilot projects to firm-wide transformation.
AI governance adoption hits 2026 inflection point amid oversight gaps
Survey data reveals rising AI adoption across enterprises in 2026, but governance maturity and accountability frameworks lag behind deployment—creating regulatory and operational risk.