Add-in DSE Brain: A Domain-Specific Expertise Architecture for Governing LLM Reasoning via a Graph-Guided SEPath Framework and Causal AI
Year: 2026
Type: preprint
Abstract: While Large Language Models (LLMs) have achieved widespread adoption across the financial sector, pure neural architectures still suffer from inherent stochasticity, creating severe bottlenecks in mission-critical, zero-tolerance operations such as corporate syndicated lending, cross-asset portfolio stress testing, and regulatory compliance. Currently, financial institutions rely heavily on costly frontier generalist LLMs, which are plagued by numerical hallucinations, black-box opacity, lack of auditability, and cloud data-leakage risks. Furthermore, because these models rely on surface-level statistical correlations, they are highly susceptible to spurious correlations and confounding biases. Conversely, on-premises compact open-weight LLMs offer superior data localization and cost efficiency, yet their constrained parameter scales prevent them from independently executing complex multi-hop financial logic and causal inference. To overcome these limitations, this study proposes a hybrid neuro-symbolic architecture that decouples natural language representation from causal-logical governance by integrating an on-premises compact open-weight LLM, Retrieval-Augmented Generation (RAG), and an add-in Domain-Specific Expertise Brain (DSE Brain). Functioning as a Cognitive Exocortex grounded in the five-layer Structured Expert Path (SEPath) control architecture, the DSE Brain upgrades static knowledge graphs into a universal, dual-regime Causal-Constrained Logic Graph (CC-LG), such as the Causal-Constrained Financial Logic Graph (CC-FLG) demonstrated in this financial setting, that distinguishes data-supported empirical variables from normative regulatory and procedural rules. Rather than forcing statistical estimation where observational data is unavailable, the system selectively invokes multi-paradigm Directed Acyclic Graph (DAG) causal discovery (PC, GES, and DirectLiNGAM), dual-track back-door and Double Machine Learning (DML) estimation, and triple-robustness refutation diagnostics exclusively to calibrate empirical causal impact weights for data-supported DEPENDS_ON dependencies, while enforcing codified prerequisite and safety assertions over normative nodes. Executed through a five-layer CoT-CR-ReAct pipeline with dual compliance gates, external deterministic solvers, and closed-loop meta-learning, the architecture actively governs latent reasoning trajectories and agentic actions. Empirical validations across global financial market series, structural prompt auditing (achieving the maximum 5.00 audit score), and longitudinal FinQA benchmarks (\approx92.0% mean accuracy) demonstrate that equipping compact open-weight LLMs with the DSE Brain transforms them into deterministic, white-box auditable inference engines that surpass frontier generalist LLMs in reasoning depth, compliance fidelity, and cost efficiency.
Author: Chou, Tsung-Nan
Institution: Chaoyang University of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237345443
Cites:
Cited by:
Related to:
