Marketing Manager
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.
A close reading of Du et al.'s ICML 2024 multiagent debate paper — which reports 14.8-point accuracy gains on arithmetic — alongside 2025 rebuttals showing equal-budget single agents match debate performance, and an analysis of why Collective Delusion (65% of debate failures) poses specific risks for AI-assisted ledger commits.
A NeurIPS 2024 Spotlight paper ablates three LLM-based time series forecasting methods — OneFitsAll, Time-LLM, and CALF — and finds that removing the language model improves accuracy in most cases, with up to a 1,383× training speedup. For finance AI applications like Beancount balance prediction, lightweight purpose-built models consistently beat repurposed LLMs.
AuditCopilot applies open-source LLMs (Mistral-8B, Gemma, Llama-3.1) to corporate journal entry fraud detection, cutting false positives from 942 to 12 — but ablation reveals the LLM functions primarily as a synthesis layer on top of Isolation Forest scores, not as an independent anomaly detector.
TAT-LLM fine-tunt LLaMA 2 7B met LoRA op financiële tabel-tekst QA-benchmarks en behaalt 64,60% EM op FinQA — waarmee het de 63,91% van GPT-4 verslaat — door redenering te ontleden in deterministische Extraheer-Redeneer-Voer-uit stappen die rekenkundige fouten elimineren.
Empirical comparison of RAG vs. unsupervised fine-tuning across 7B-parameter LLMs shows RAG achieves 0.875+ accuracy on post-cutoff facts while fine-tuning plateaus at 0.504 — with direct implications for Beancount agent design and any system requiring frequent knowledge updates.
IRCoT interleaves BM25 retrieval with each step of a chain-of-thought reasoning loop, achieving +11.3 retrieval recall and +7.1 F1 on HotpotQA over one-step RAG — and shows a 3B model can beat GPT-3 175B when retrieval strategy is right.
FLARE (EMNLP 2023) improves on standard RAG by triggering retrieval mid-generation using token-probability confidence thresholds, reaching 51.0 EM on 2WikiMultihopQA versus 39.4 for single-retrieval — but calibration failures in instruction-tuned chat models limit its reliability for production finance agents.
Lewis et al.'s NeurIPS 2020 paper introduced the hybrid RAG architecture—a BART-large generator paired with a FAISS-indexed retriever over 21 million Wikipedia passages—achieving 44.5 EM on Natural Questions and establishing the parametric/non-parametric split that now underlies most production AI systems. This review covers RAG-Sequence vs. RAG-Token trade-offs, the retrieval collapse failure mode, and what stale indexes mean for financial AI built on append-only Beancount ledgers.
MultiHiertt (ACL 2022) introduces 10,440 QA pairs from real financial reports averaging 3.89 hierarchical tables each; state-of-the-art models score 38% F1 versus 87% for humans, with a 15-point penalty for cross-table questions — quantifying the retrieval gap finance AI must close.
ConvFinQA (EMNLP 2022) extends FinQA into multi-turn conversation over S&P 500 earnings reports, finding that the best fine-tuned model achieves 68.9% execution accuracy versus 89.4% for human experts—and drops to 52.4% on hybrid multi-aspect conversations where models must carry numerical context across different financial topics.
TAT-QA is a 16,552-question benchmark over hybrid table-plus-text financial report contexts that showed evidence grounding — not arithmetic — is the core bottleneck in finance AI; by 2024, fine-tuned 7B LLMs reached 83% F1, closing most of the gap against a 91% human ceiling.
