Selected work

Representative Papers

PetroBench: A Benchmark for Large Language Models in Petroleum Engineering

May 27, 2026

This study addresses the absence of domain-specific evaluation benchmarks for large language models (LLMs) in petroleum engineering by introducing the first standardized assessment framework encompassing production, reservoir, and drilling engineering, comprising 1,200 multi-format questions. Data quality is ensured through a rigorous three-stage pipeline involving expert review, preprocessing, and quality filtering, followed by validation across multiple models. Systematic evaluations of leading Chinese and English LLMs are conducted under a unified API environment. Results reveal that models perform better on subjective than objective questions, achieving peak accuracies of 65.3% and 74.3% on multiple-choice and true/false items, respectively. Models such as Gemini-1.5-Pro attain overall scores of 72%–74%, with Chinese models excelling in multiple-choice tasks and international models showing slight advantages in short-answer responses. The benchmark demonstrates strong domain relevance, high discriminative power, and reproducibility.

0 citationsRead paper

Hallucination as a Computational Boundary: A Hierarchy of Inevitability and the Oracle Escape

Aug 10, 2025

Hallucination in large language models (LLMs) severely undermines their reliable deployment. Method: This paper establishes the first formal theoretical framework demonstrating that hallucination is fundamentally unavoidable—arising from computational inevitabilities rooted in diagonalization, uncomputability, and information-theoretic limits. We introduce the “Learner Pumping Lemma” and integrate probabilistic Turing machine modeling, neural game theory, and computational jump theory to rigorously analyze retrieval-augmented generation (RAG) and continual learning. Contributions/Results: First, we provide the first rigorous formal foundation for RAG. Second, we model RAG as an external oracle mechanism, enabling *absolute* hallucination avoidance under idealized assumptions. Third, we formalize continual learning as oracle internalization—the progressive incorporation of external knowledge into model parameters. Collectively, these results reveal two fundamental, complementary pathways for hallucination mitigation: *exogenous retrieval* (leveraging external knowledge sources) and *endogenous knowledge integration* (embedding verified knowledge into model representations).

0 citationsRead paper

Interpretable Style Takagi-Sugeno-Kang Fuzzy Clustering

Apr 07, 2025

Existing clustering methods generally lack interpretability and neglect intrinsic inter-group style heterogeneity and intra-group homogeneity. To address this, we propose an interpretable stylized TSK fuzzy clustering method: it unsupervisedly generates clusters represented by rule consequent vectors, and explicitly models intra-group homogeneity and inter-group stylistic discrepancies via a learnable style matrix—thereby achieving dual interpretability of cluster structure and decision logic. This work is the first to deeply integrate style modeling with Takagi–Sugeno–Kang (TSK) fuzzy inference into unsupervised clustering, enabling adaptive identification of both explicit and implicit data styles. Extensive experiments on diverse benchmark datasets demonstrate that our method significantly outperforms state-of-the-art clustering algorithms, especially in style-sensitive tasks. The source code is publicly available.

0 citationsRead paper

Pig behavior dataset and Spatial-temporal perception and enhancement networks based on the attention mechanism for pig behavior recognition

Mar 12, 2025

To address the longstanding absence of publicly available, fine-grained video datasets for porcine behavior recognition, this study introduces the first open-source video dataset comprising 13 welfare-critical pig behaviors. We further propose ST-ANet, an attention-guided spatiotemporal awareness and enhancement network featuring a two-stage architecture: Stage I localizes behavior-relevant regions and models individual and interactive dynamics via spatiotemporal graph convolution; Stage II incorporates feature recalibration and long-range temporal modeling to strengthen spatiotemporal dependencies. Evaluated on our curated dataset, ST-ANet achieves a mean Average Precision (mAP) of 75.92%, outperforming the best conventional method by 8.17 percentage points. The framework significantly improves accuracy in individual behavior recognition and demonstrates enhanced generalizability across diverse farming scenarios.

0 citationsRead paper

Recent publications

Latest Papers

PetroBench: A Benchmark for Large Language Models in Petroleum Engineering

May 27, 2026

This study addresses the absence of domain-specific evaluation benchmarks for large language models (LLMs) in petroleum engineering by introducing the first standardized assessment framework encompassing production, reservoir, and drilling engineering, comprising 1,200 multi-format questions. Data quality is ensured through a rigorous three-stage pipeline involving expert review, preprocessing, and quality filtering, followed by validation across multiple models. Systematic evaluations of leading Chinese and English LLMs are conducted under a unified API environment. Results reveal that models perform better on subjective than objective questions, achieving peak accuracies of 65.3% and 74.3% on multiple-choice and true/false items, respectively. Models such as Gemini-1.5-Pro attain overall scores of 72%–74%, with Chinese models excelling in multiple-choice tasks and international models showing slight advantages in short-answer responses. The benchmark demonstrates strong domain relevance, high discriminative power, and reproducibility.

0 citationsRead paper

Hallucination as a Computational Boundary: A Hierarchy of Inevitability and the Oracle Escape

Aug 10, 2025

Hallucination in large language models (LLMs) severely undermines their reliable deployment. Method: This paper establishes the first formal theoretical framework demonstrating that hallucination is fundamentally unavoidable—arising from computational inevitabilities rooted in diagonalization, uncomputability, and information-theoretic limits. We introduce the “Learner Pumping Lemma” and integrate probabilistic Turing machine modeling, neural game theory, and computational jump theory to rigorously analyze retrieval-augmented generation (RAG) and continual learning. Contributions/Results: First, we provide the first rigorous formal foundation for RAG. Second, we model RAG as an external oracle mechanism, enabling *absolute* hallucination avoidance under idealized assumptions. Third, we formalize continual learning as oracle internalization—the progressive incorporation of external knowledge into model parameters. Collectively, these results reveal two fundamental, complementary pathways for hallucination mitigation: *exogenous retrieval* (leveraging external knowledge sources) and *endogenous knowledge integration* (embedding verified knowledge into model representations).

0 citationsRead paper

Interpretable Style Takagi-Sugeno-Kang Fuzzy Clustering

Apr 07, 2025

Existing clustering methods generally lack interpretability and neglect intrinsic inter-group style heterogeneity and intra-group homogeneity. To address this, we propose an interpretable stylized TSK fuzzy clustering method: it unsupervisedly generates clusters represented by rule consequent vectors, and explicitly models intra-group homogeneity and inter-group stylistic discrepancies via a learnable style matrix—thereby achieving dual interpretability of cluster structure and decision logic. This work is the first to deeply integrate style modeling with Takagi–Sugeno–Kang (TSK) fuzzy inference into unsupervised clustering, enabling adaptive identification of both explicit and implicit data styles. Extensive experiments on diverse benchmark datasets demonstrate that our method significantly outperforms state-of-the-art clustering algorithms, especially in style-sensitive tasks. The source code is publicly available.

0 citationsRead paper

Pig behavior dataset and Spatial-temporal perception and enhancement networks based on the attention mechanism for pig behavior recognition

Mar 12, 2025

To address the longstanding absence of publicly available, fine-grained video datasets for porcine behavior recognition, this study introduces the first open-source video dataset comprising 13 welfare-critical pig behaviors. We further propose ST-ANet, an attention-guided spatiotemporal awareness and enhancement network featuring a two-stage architecture: Stage I localizes behavior-relevant regions and models individual and interactive dynamics via spatiotemporal graph convolution; Stage II incorporates feature recalibration and long-range temporal modeling to strengthen spatiotemporal dependencies. Evaluated on our curated dataset, ST-ANet achieves a mean Average Precision (mAP) of 75.92%, outperforming the best conventional method by 8.17 percentage points. The framework significantly improves accuracy in individual behavior recognition and demonstrates enhanced generalizability across diverse farming scenarios.

0 citationsRead paper