#trust
Reliability, calibration, and hallucination in financial AI systems
A systematic survey of LLM confidence estimation and calibration methods—white-box logit approaches, consistency-based SelfCheckGPT, and semantic entropy—reveals that verbalized confidence scores from GPT-4 achieve only ~62.7% AUROC, barely above chance, with direct implications for deploying uncertainty-aware agents in finance and accounting.
ReDAct runs a small model by default and escalates to an expensive model only when token-level perplexity signals uncertainty, achieving 64% cost savings over GPT-5.2-only while matching or exceeding its accuracy — a directly applicable pattern for Beancount transaction-categorization agents.
CMU and NC State researchers propose using System-Theoretic Process Analysis (STPA) and a capability-enhanced Model Context Protocol to derive formal safety specifications for LLM agent tool use, with Alloy-based verification demonstrating absence of unsafe flows in a calendar scheduling case study.
AGrail (ACL 2025) introduces a two-LLM cooperative guardrail that adapts safety checks at inference time via test-time adaptation, achieving 0% prompt injection attack success and 95.6% benign action preservation on Safe-OS — compared to GuardAgent and LLaMA-Guard blocking up to 49.2% of legitimate actions.
ShieldAgent (ICML 2025) replaces LLM-based guardrails with probabilistic rule circuits built on Markov Logic Networks, achieving 90.4% accuracy on agent attacks with 64.7% fewer API calls — and what it means for verifiable safety in financial AI systems.
GuardAgent (ICML 2025) places a separate LLM agent between a target agent and its environment, verifying every proposed action by generating and running Python code — achieving 98.7% policy enforcement accuracy while preserving 100% task completion, versus 81% accuracy and 29–71% task failure for prompt-embedded safety rules.
Huang et al. (ICLR 2024): intrinsic self-correction drops GPT-4 from 95.5% to 91.5% on GSM8K—ledger agents need external validators, not self-review.
PHANTOM (NeurIPS 2025) is the first benchmark to measure LLM hallucination detection on real SEC filings across context lengths up to 30,000 tokens. Qwen3-30B-A3B-Thinking leads with F1=0.882; 7B models score near random guessing — with direct implications for autonomous accounting agents.
