arXiv subjects

Pierre Jinghong Liang

Publications and source records attributed to Pierre Jinghong Liang.

4 recordsLinked to original sources

arXiv2026

FinGREAT: Evaluating LLMs on Financial Graph-guided Retrieval And Typing

Financial disclosures express the underlying relationships of double-entry accounting through unstructured text. While financial LLM benchmarks predominantly emphasize question answering and classification, schema-guided evidence discovery remains relatively under-explored. We investigate whether open-weight LLMs can extract and classify evidence corresponding to predefined double-entry accounting relationships underlying the text. Using a financial accounting graph comprising 11 nodes (accounts) and 11 edges (transaction types), we evaluate mid-sized models from 5 open-weight families across 1,300 SEC 10-K Item 7 filings spanning 26 years, comparing their extraction performance against three proprietary models. Results reveal substantial differences in extraction volume, edge distributions, and cross-model agreement. Building on this study, we introduce a human-annotated relation classification benchmark comprising 483 passages across three financial accounting categories, evaluating 34 open-weight and 10 proprietary models. The strongest open-weight and proprietary models achieve 84.1\% and 92.8\% accuracy, respectively, with considerable variation across accounting categories. Our findings highlight the importance of evaluating evidence discovery and relation understanding separately, while investigating locally deployable models for scalable, privacy-sensitive financial analysis.

cs.LG↗

arXiv2026

Structured Prompt Optimization Meets Reinforcement Learning for Global and Local Interpretability over Complex Text

LLMs have advanced text classification, yet existing paradigms face a trade-off: supervised (label only) fine-tuning is scalable but offers limited reasoning on complex text and lacks broader model transparency, while discrete prompt optimization offers human-readable instructions but struggles with performance and scalability. We introduce eXTC (eXplainable Text Classifier) with three progressive stages: (1) learning a Standard Operating Procedure (SOP, or rulebook) in natural language via a new Structured Prompt Optimization algorithm; (2) SOP-grounded reasoning distillation from a large teacher LLM into a compact LM; and (3) expanding reasoning capabilities beyond the initial SOP via reinforcement learning. This design enables eXTC to provide (i) fast inference via a compact LM, with (ii) inference-time local reasoning traces, alongside a global, modular explanation of its learned domain rules, while (iii) significantly outperforming existing paradigms across diverse benchmarks in both classification performance and explanation quality, with stage-by-stage gains.

cs.CL↗

arXiv2023

ADAMM: Anomaly Detection of Attributed Multi-graphs with Metadata: A Unified Neural Network Approach

Given a complex graph database of node- and edge-attributed multi-graphs as well as associated metadata for each graph, how can we spot the anomalous instances? Many real-world problems can be cast as graph inference tasks where the graph representation could capture complex relational phenomena (e.g., transactions among financial accounts in a journal entry), along with metadata reflecting tabular features (e.g. approver, effective date, etc.). While numerous anomaly detectors based on Graph Neural Networks (GNNs) have been proposed, none are capable of directly handling directed graphs with multi-edges and self-loops. Furthermore, the simultaneous handling of relational and tabular features remains an unexplored area. In this work we propose ADAMM, a novel graph neural network model that handles directed multi-graphs, providing a unified end-to-end architecture that fuses metadata and graph-level representation learning through an unsupervised anomaly detection objective. Experiments on datasets from two different domains, namely, general-ledger journal entries from different firms (accounting) as well as human GPS trajectories from thousands of individuals (urban mobility) validate ADAMM's generality and detection effectiveness of expert-guided and ground-truth anomalies. Notably, ADAMM outperforms existing baselines that handle the two data modalities (graph and metadata) separately with post hoc synthesis efforts.

cs.LG↗

arXiv2020

AutoAudit: Mining Accounting and Time-Evolving Graphs

How can we spot money laundering in large-scale graph-like accounting datasets? How to identify the most suspicious period in a time-evolving accounting graph? What kind of accounts and events should practitioners prioritize under time constraints? To tackle these crucial challenges in accounting and auditing tasks, we propose a flexible system called AutoAudit, which can be valuable for auditors and risk management professionals. To sum up, there are four major advantages of the proposed system: (a) "Smurfing" Detection, spots nearly 100% of injected money laundering transactions automatically in real-world datasets. (b) Attention Routing, attends to the most suspicious part of time-evolving graphs and provides an intuitive interpretation. (c) Insight Discovery, identifies similar month-pair patterns proved by "success stories" and patterns following Power Laws in log-logistic scales. (d) Scalability and Generality, ensures AutoAudit scales linearly and can be easily extended to other real-world graph datasets. Experiments on various real-world datasets illustrate the effectiveness of our method. To facilitate reproducibility and accessibility, we make the code, figure, and results public at https://github.com/mengchillee/AutoAudit.

cs.SI↗