Paper: arXiv 2512.14744
Authors: Adewale Akinfaderin, Shreyas Subramanian
Abstract
Financial AI systems suffer from a critical blind spot: while Retrieval-Augmented Generation (RAG) excels at finding relevant documents, language models still generate calculation errors and regulatory violations during reasoning, even with perfect retrieval. This paper introduces VERAFI (Verified Agentic Financial Intelligence), an agentic framework with neurosymbolic policy generation for verified financial intelligence. VERAFI combines state-of-the-art dense retrieval and cross-encoder reranking with financial tool-enabled agents and automated reasoning policies covering GAAP compliance, SEC requirements, and mathematical validation. Our comprehensive evaluation on FinanceBench demonstrates remarkable improvements: while traditional dense retrieval with reranking achieves only 52.4% factual correctness, VERAFI’s integrated approach reaches 94.7%, an 81% relative improvement. The neurosymbolic policy layer alone contributes a 4.3 percentage point gain over pure agentic processing, specifically targeting persistent mathematical and logical errors. By integrating financial domain expertise directly into the reasoning process, VERAFI offers a practical pathway toward trustworthy financial AI that meets the stringent accuracy demands of regulatory compliance, investment decisions, and risk management.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper introduces a novel neurosymbolic policy layer using formal SMT-lib specifications for regulatory and mathematical validation, which involves advanced mathematical formalization. It also provides comprehensive empirical evaluation on FinanceBench with specific metrics (94.7% factual correctness, 81% relative improvement), demonstrating backtest-ready rigor.
Research Flowchart
flowchart TD
A["Research Goal: Verify Financial AI"] --> B["Data & Framework"]
B --> C["Method: Neurosymbolic Policy"]
C --> D["Process: Agentic Tools & RAG"]
D --> E["Outcome: 94.7% Accuracy"]
subgraph B ["Inputs"]
B1["FinanceBench Dataset"]
B2["GAAP/SEC Regulations"]
end
subgraph D ["Computation"]
D1["Dense Retrieval"]
D2["Cross-Encoder Reranking"]
D3["Tool-Enabled Reasoning"]
end
subgraph E ["Results"]
E1["81% Relative Improvement"]
E2["Regulatory Compliance"]
E3["Mathematical Verification"]
end
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes VERAFI, a neurosymbolic agentic framework that integrates retrieval, tool-use, and policy‑guided generation for verified financial intelligence.
Source sentence
This paper introduces VERAFI (Verified Agentic Financial Intelligence), an agentic framework with neurosymbolic policy generation for verified financial intelligence.
Assumes
No explicit assumptions found in the text.
Data
- FinanceBench subset benchmark public
Source sentence
we construct our evaluation dataset using a targeted subset of companies from FinanceBench (Islam et al. 2023) and ConvFinQA (Chen et al. 2022). Our dataset focuses on four major corporations: American Water Works (2015-2018 filings), AMD (2022 filings), American Express (2022 filings), and Boeing (2022 filings).
- ConvFinQA benchmark public
Source sentence
we construct our evaluation dataset using a targeted subset of companies from FinanceBench (Islam et al. 2023) and ConvFinQA (Chen et al. 2022).
Costs
Not applicable
Finds 4
- VERAFI achieves 94.7% factual correctness and 96.4% completeness, an 81% relative improvement over traditional dense retrieval with reranking. performance
Source sentence
Our best‑performing configuration, VERAFI with neurosymbolic validation, achieves 94.7% factual correctness and 96.4% completeness, substantially outperforming all baseline methods
- The neurosymbolic policy layer adds a 4.3 percentage‑point gain over the pure agentic baseline. performance
Source sentence
The neurosymbolic policy layer alone contributes a 4.3 percentage point gain over pure agentic processing, specifically targeting persistent mathematical and logical errors.
- Dense retrieval combined with cross‑encoder reranking (Dense+Rerank) attains 66.7% recall, the best retrieval performance among tested methods. performance
Source sentence
Dense+Rerank achieves best performance, dominating across all metrics with 66.7% recall, successfully retrieving relevant documents for approximately two‑thirds of financial queries.
- Agentic processing with web search raises factual correctness from 52.4% (Dense+Reranker) to 90.4% before neurosymbolic validation. performance
Source sentence
factual correctness increasing from 52.4% (Dense + Reranker) to 90.4% with agentic processing and web search, and reaching 94.7% with neurosymbolic validation.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 13 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (32)
- Retrieval-augmented generation for knowledge-intensive nlp tasks (2020)
- Dense Passage Retrieval for Open-Domain Question Answering. (2020)
- Latent Retrieval for Weakly Supervised Open Domain Question Answering (2019)
- Retrieval augmented language model pre-training (2020)
- Passage Re-ranking with BERT (2019) — link
- Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering (2021)
- Improving language models by retrieving from trillions of tokens (2022)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (2024)
and 24 more
- Webgpt: Browser-assisted question-answering with human feedback (2021) — link
- Active retrieval augmented generation (2023)
- Agentic retrieval-augmented generation: A survey on agentic rag (2025) — link
- Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries (2024) — link
- Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions (2023)
- FinQA: A Dataset of Numerical Reasoning over Financial Data (2021)
- ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering (2022)
- FinanceBench: A New Benchmark for Financial Question Answering (2023)
- Financial analysis: Intelligent financial data analysis system based on llm-rag — → 2504.06279
- Enhancing Financial RAG with Agentic AI and Multi-HyDE: A Novel Approach to Knowledge Retrieval and Hallucination Reduction (2025) — link
- Finance Agent Benchmark: Benchmarking LLMs on Real-world Financial Research Tasks (2025) — link
- SymRAG: Efficient Neuro-Symbolic Retrieval Through Adaptive Query Routing (2025)
- Scaling Scientific Knowledge Discovery with Neuro-Symbolic AI and Large Language Models (2024)
- Sound and Complete Neurosymbolic Reasoning with LLM-Grounded Interpretations (2025)
- A Neuro-Symbolic Approach to Multi-Agent RL for Interpretability and Probabilistic Decision Making (2024) — link
- Enhancing Large Language Models with Neurosymbolic Reasoning for Multilingual Tasks (2025) — link
- Minimize generative AI hallucinations with Amazon Bedrock Automated Reasoning checks (2025) — link
- Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models (2025) — link
- jina-reranker-v3: Last but Not Late Interaction for Document Reranking (2025) — link
- The smt-lib standard: Version 2.0 (2010)
- Judging llm-as-a-judge with mt-bench and chatbot arena (2023)
- G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment (2023)
- LLM-as-a-judge on Amazon Bedrock Model Evaluation (2025) — link
- A Neurosymbolic Approach to Natural Language Formalization and Verification (2025) — link
1 of 32 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.