arXiv subjects

Xiaolong Liang

Publications and source records attributed to Xiaolong Liang.

13 recordsLinked to original sources

arXiv2026

FinVault: Benchmarking Financial Agent Safety in Execution-Grounded Environments

Financial agents powered by large language models (LLMs) are increasingly deployed for investment analysis, risk assessment, and automated decision-making, where their abilities to plan, invoke tools, and manipulate mutable state introduce new security risks in high-stakes and highly regulated financial environments. However, existing safety evaluations largely focus on language-model-level content compliance or abstract agent settings, failing to capture execution-grounded risks arising from real operational workflows and state-changing actions. To bridge this gap, we propose FinVault, the first execution-grounded security benchmark for financial agents, comprising 31 regulatory case-driven sandbox scenarios with state-writable databases and explicit compliance constraints, together with 107 real-world vulnerabilities and 963 test cases that systematically cover prompt injection, jailbreaking, financially adapted attacks, as well as benign inputs for false-positive evaluation. Experimental results reveal that existing defense mechanisms remain ineffective in realistic financial agent settings, with average attack success rates (ASR) still reaching up to 50.0\% on state-of-the-art models and remaining non-negligible even for the most robust systems (ASR 6.7\%), highlighting the limited transferability of current safety designs and the need for stronger financial-specific defenses. Our code can be found at https://github.com/aifinlab/FinVault.

cs.CR↗

arXiv2026

Agree on the Model, Verify the Inference: GKR Protocols for HND-Based Transformer Inference

Outsourced Transformer inference exposes clients to model substitution and incomplete execution, while direct replay removes the computational benefit of delegation. We present GKR-HND, a registered-model protocol for verifying the polynomial backbone of Homomorphic--Nonhomomorphic Decomposition Transformers. The retained verifier checks the GKR transcript and registered-weight openings, but delegates expensive public evaluations to an assigned computation worker. Assuming an honest retained verifier and prover--worker non-collusion, the verifier accepts only when the worker's signed, request-bound response agrees with the proof claims. Experiments with pretrained HND models validate the proof path and the delegated public computation without dense-matrix replay.

cs.LG↗

arXiv2026

Epistemic Skills: Reasoning about Knowledge and Oblivion

This paper presents a class of epistemic logics that captures the dynamics of acquiring knowledge and descending into oblivion, while incorporating concepts of group knowledge. The approach is grounded in a system of weighted models, introducing an ``epistemic skills'' metric to represent the epistemic capacities tied to knowledge updates. Within this framework, knowledge acquisition is modeled as a process of upskilling, whereas oblivion is represented as a consequence of downskilling. The framework further enables exploration of ``knowability'' and ``forgettability,'' defined as the potential to gain knowledge through upskilling and to lapse into oblivion through downskilling, respectively. Additionally, it supports a detailed analysis of the distinctions between epistemic de re and de dicto expressions. The computational complexity of the model checking and satisfiability problems is examined, offering insights into their theoretical foundations and practical implications.

cs.AI↗

arXiv2026

UniFinEval: Towards Unified Evaluation of Financial Multimodal Models across Text, Images and Videos

Multimodal large language models are playing an increasingly significant role in empowering the financial domain, however, the challenges they face, such as multimodal and high-density information and cross-modal multi-hop reasoning, go beyond the evaluation scope of existing multimodal benchmarks. To address this gap, we propose UniFinEval, the first unified multimodal benchmark designed for high-information-density financial environments, covering text, images, and videos. UniFinEval systematically constructs five core financial scenarios grounded in real-world financial systems: Financial Statement Auditing, Company Fundamental Reasoning, Industry Trend Insights, Financial Risk Sensing, and Asset Allocation Analysis. We manually construct a high-quality dataset consisting of 3,767 question-answer pairs in both chinese and english and systematically evaluate 10 mainstream MLLMs under Zero-Shot and CoT settings. Results show that Gemini-3-pro-preview achieves the best overall performance, yet still exhibits a substantial gap compared to financial experts. Further error analysis reveals systematic deficiencies in current models. UniFinEval aims to provide a systematic assessment of MLLMs' capabilities in fine-grained, high-information-density financial environments, thereby enhancing the robustness of MLLMs applications in real-world financial scenarios. Data and code are available at https://github.com/aifinlab/UniFinEval.

q-fin.GN↗

arXiv2025

Sequential-NIAH: A Needle-In-A-Haystack Benchmark for Extracting Sequential Needles from Long Contexts

Evaluating the ability of large language models (LLMs) to process lengthy contexts is critical, especially for retrieving query-relevant information embedded within them. We introduce Sequential-NIAH, a benchmark specifically designed to evaluate the capability of LLMs to extract sequential information items (known as \emph{needles}) from long contexts. The benchmark includes three needle generation pipelines: synthetic-temporal, real-temporal, and real-logical orders, with context lengths ranging from 8K to 128K, which comprises 14,000 samples (2,000 for testing). To facilitate the evaluation of this benchmark, we trained an evaluation model that assesses the correctness of LLM responses by comparing their completeness and sequential consistency against the ground truth, which provides a more reliable evaluation metric than GPT-4 or Claude. We conducted experiments on six well-known LLMs, revealing that even the best-performing model achieved a maximum accuracy of only 63.50% on test set of this benchmark. Further analysis highlights the growing challenges posed by increasing the context length or the number of needles, underscoring substantial room for improvement of LLMs. Additionally, noise analysis validates the reliability and challenge of the benchmark, making Sequential-NIAH an important reference for advancing research on long text information extraction capabilities of LLMs.

cs.CL↗

arXiv2025

Use of electrical resistivity tomography to map the tree roots

An efficient advanced numerical model for mapping the distribution of the buried tree roots is presented. It not only simplify the complicate root branches to an easy manipulated model, but also grasp the main structure of tree roots ignoring the unnecessary minutiae, and thus provide an intuitive impression of subsurface invisible anomalies. The processing model is combined with an adaptive finite element method, which can automatically generate unstructured triangular meshes during the process of discretization, which also enable user to specifically set the resistivity along each part of tree roots.

physics.geo-ph↗

arXiv2025

Observational evidence of anisotropic changes apparent resistivity before strong earthquakes

Using a method based on normalized monthly variation rate, we studied resistivity data of seven observation stations before the events in the epicenter areas of two strong earthquakes. The relationship between variation of anisotropic apparent resistivity and the azimuth of the maximum principal stress is analyzed. The study shows that significant apparent resistivity variation occurs in the direction that is perpendicular to the azimuth of the maximum principal stress while only small fluctuation are recorded in the direction of the maximum principal stress. We surmise that the variation of anisotropic resistivity occurs in the late stage of the development of a strong earthquake, which can be observed in the epicenter area. If the density of the observation stations is increased and the direction of the observed resistivity is right, the epicenter of an earthquake location may be estimated by the observed resistivity anomaly.

physics.geo-ph↗

arXiv2025

Use of Ground Penetrating Radar to Map the Tree Roots

Tree roots can support and transmit nutrients for trees healthy growth aboveground, which greatly improve trees productivity and have significant effect on maintaining the normal operation of ecosystem. In order to map the tree roots more efficiently and effectively, the nondestructive ground penetrating radar is introduced into this area. The construction of tree roots model mainly conducted by the profile matrix which stored electromagnetic parameters of tree roots, ground penetrating radar set the normalized first derivative Blackman-Harris window function as the source pulse. Two-way travel time, the electromagnetic pulses arriving at root zone and then reflected back to the receive antenna, which can be calculated by two-dimensional Finite-Difference Time-Domain. Finally synthesized the common-offset reflection data that extracted from the output multi-offset data cube as radargrams which contain the information about buried tree roots. The results turned out that through interaction between electromagnetic pulse and underground anomalies, the distribution information related subsurface buried tree roots can be observed accurately from radargrams, in addition to the intermediate section shielded by tree roots barrier, the dipping boundary between clay layer and bedrock layer is clear enough to be noticed. With the increase of radar frequency, the electromagnetic pulse meet severe attenuation accompanied by the detection depth decrease, thus the texture in radargram gradually blurred. These relatively accurate roots outline, calculated by numerical simulation, showed that the application of ground penetrating radar in tree roots detection can significantly improve resolution of roots which stretched in the vertical direction.

physics.geo-ph↗

arXiv2024

FinEval: A Chinese Financial Domain Knowledge Evaluation Benchmark for Large Language Models

Large language models have demonstrated outstanding performance in various natural language processing tasks, but their security capabilities in the financial domain have not been explored, and their performance on complex tasks like financial agent remains unknown. This paper presents FinEval, a benchmark designed to evaluate LLMs' financial domain knowledge and practical abilities. The dataset contains 8,351 questions categorized into four different key areas: Financial Academic Knowledge, Financial Industry Knowledge, Financial Security Knowledge, and Financial Agent. Financial Academic Knowledge comprises 4,661 multiple-choice questions spanning 34 subjects such as finance and economics. Financial Industry Knowledge contains 1,434 questions covering practical scenarios like investment research. Financial Security Knowledge assesses models through 1,640 questions on topics like application security and cryptography. Financial Agent evaluates tool usage and complex reasoning with 616 questions. FinEval has multiple evaluation settings, including zero-shot, five-shot with chain-of-thought, and assesses model performance using objective and subjective criteria. Our results show that Claude 3.5-Sonnet achieves the highest weighted average score of 72.9 across all financial domain categories under zero-shot setting. Our work provides a comprehensive benchmark closely aligned with Chinese financial domain.

cs.CL↗

arXiv2024

Epistemic Skills: Logical Dynamics of Knowing and Forgetting

We present a type of epistemic logics that encapsulates both the dynamics of acquiring knowledge (knowing) and losing information (forgetting), alongside the integration of group knowledge concepts. Our approach is underpinned by a system of weighted models, which introduces an "epistemic skills" metric to effectively represent the epistemic abilities associated with knowledge update. In this framework, the acquisition of knowledge is modeled as a result of upskilling, whereas forgetting is by downskilling. Additionally, our framework allows us to explore the concept of "knowability," which can be defined as the potential to acquire knowledge through upskilling, and facilitates a nuanced understanding of the distinctions between epistemic de re and de dicto expressions. We study the computational complexity of model checking problems for these logics, providing insights into both the theoretical underpinnings and practical implications of our approach.

cs.LO↗

arXiv2024

Field Knowledge as a Dual to Distributed Knowledge: A Characterization by Weighted Modal Logic

The study of group knowledge concepts such as mutual, common, and distributed knowledge is well established within the discipline of epistemic logic. In this work, we incorporate epistemic abilities of agents to refine the formal definition of distributed knowledge and introduce a formal characterization of field knowledge. We propose that field knowledge serves as a dual to distributed knowledge. Our approach utilizes epistemic logics with various group knowledge constructs, interpreted through weighted models. We delve into the eight logics that stem from these considerations, explore their relative expressivity and develop sound and complete axiomatic systems.

cs.LO↗

arXiv2023

Epistemic Logic over Similarity Graphs: Common, Distributed and Mutual Knowledge

In this paper, we delve into the study of epistemic logics, interpreted through similarity models based on weighted graphs. We explore eight languages that extend the traditional epistemic language by incorporating modalities of common, distributed, and mutual knowledge. The concept of individual knowledge is redefined under these similarity models. It is no longer just a matter of personal knowledge, but is now enriched and understood as knowledge under the individual's epistemic ability. Common knowledge is presented as higher-order knowledge that is universally known to any degree, a definition that aligns with existing literature. We reframe distributed knowledge as a form of knowledge acquired by collectively leveraging the abilities of a group of agents. In contrast, mutual knowledge is defined as the knowledge obtained through the shared abilities of a group. We then focus on the resulting logics, examining their relative expressivity, semantic correspondence to the classical epistemic logic, proof systems and the computational complexity associated with the model checking problem and the satisfiability/validity problem. This paper offers significant insights into the logical analysis and understanding of these enriched forms of knowledge, contributing to the broader discourse on epistemic logic.

cs.LO↗

arXiv2019

Adversarial Defense by Suppressing High-frequency Components

Recent works show that deep neural networks trained on image classification dataset bias towards textures. Those models are easily fooled by applying small high-frequency perturbations to clean images. In this paper, we learn robust image classification models by removing high-frequency components. Specifically, we develop a differentiable high-frequency suppression module based on discrete Fourier transform (DFT). Combining with adversarial training, we won the 5th place in the IJCAI-2019 Alibaba Adversarial AI Challenge. Our code is available online.

cs.CV↗