Selected work

Representative Papers

Large Language Models for Programming: Actually Fixing or Reimplementing Incorrect Code?

Sep 24, 2026

This study investigates whether large language models (LLMs) tend to apply localized patches or complete rewrites when repairing code, and how their behavior diverges from that of human developers. To this end, the authors construct an evaluation benchmark based on Codeforces and present the first quantitative analysis measuring the deviation of LLM repair behaviors from human-generated patches. Furthermore, they conduct automated comparative evaluations across GPT-series models using text similarity algorithms. The findings reveal that LLMs inherently violate the principle of minimal modification, exhibiting a strong preference for reimplementation over targeted patching. Notably, the study demonstrates that generating solutions from scratch outperforms repairing erroneous code. These insights provide critical implications for the design of AI-assisted debugging tools, highlighting fundamental discrepancies between machine-driven and human-driven code repair strategies.

0 citationsRead paper

On the Triangle Inequality for the Jaccard Distance in Arbitrary Lattices

Aug 18, 2026

本文解决了在任意格上推广Jaccard距离的问题，通过证明当估值满足特定条件时Jaccard距离满足三角不等式，并探讨了其在量子信息理论等领域的应用。

0 citationsRead paper

Foundations of MT-PDCL: Measure-Theoretic Probabilistic Definite Clause Logic

Aug 13, 2026

Traditional probabilistic logic programming is constrained by discrete propositionalization, hindering precise reasoning over continuous domains. This work proposes the MT-PDCL framework, which introduces measure theory into probabilistic logic programming for the first time. By leveraging standard Borel σ-algebras and Lebesgue integration, MT-PDCL defines logical variables directly over continuous measurable spaces and establishes a semantics based on continuous probability distributions along with a continuous immediate consequence operator. The approach preserves a purely declarative syntax while enabling exact, algebraic, and structurally differentiable probabilistic inference over continuous domains. Centered on continuous integration rather than discrete combinatorial enumeration, the framework effectively circumvents the limitations of finite domains and supports efficient, exact inference with continuous priors and observations.

0 citationsRead paper

The DIME Architecture: A Unified Operational Algorithm for Neural Representation, Dynamics, Control and Integration

Mar 07, 2026

Current neuroscience lacks a unified computational architecture that integrates core cognitive functions such as perception, memory, prediction, value assessment, and consciousness. This work proposes DIME (Detect-Integrate-Mark-Execute), an operational neural architecture grounded in a four-stage cyclic mechanism that unifies neural representation, dynamic evolution, control, and multiscale integration through the interplay of memory traces, execution trajectories, a marking system, and hyper-memory traces. For the first time, DIME incorporates perception, memory, valuation, and conscious access within a single computational framework, leveraging abstract modeling, recursive processing, neuromodulation, and large-scale network integration. The architecture aligns with empirical findings including hippocampal indexing, cortical replay, and the global workspace theory, while offering a novel, interpretable, and scalable template for next-generation cognitive architectures in artificial intelligence and robotics.

0 citationsRead paper

LoopDB: A Loop Closure Dataset for Large Scale Simultaneous Localization and Mapping

Jun 07, 2025

Existing loop closure detection benchmarks suffer from limited scene diversity, insufficient scale, and sparse ground-truth pose annotations, hindering comprehensive evaluation of SLAM algorithms. To address this, we introduce LoopDB—the first open-source, multi-scene, high-resolution image-sequence benchmark specifically designed for SLAM loop closure detection. LoopDB encompasses six distinct environments (e.g., park, indoor, parking lot, and object close-ups), comprising over 1,000 images organized into 5-frame continuous sequences, with dense, frame-wise relative pose ground truth derived from high-precision camera capture and geometric calibration. The dataset is compatible with mainstream deep learning architectures, including CNNs and Vision Transformers. LoopDB significantly enhances the evaluability of generalization and robustness of loop closure methods under complex, realistic conditions. It is publicly released and has already been adopted by multiple SLAM research groups.

0 citationsRead paper

Recent publications

Latest Papers

Large Language Models for Programming: Actually Fixing or Reimplementing Incorrect Code?

Sep 24, 2026

This study investigates whether large language models (LLMs) tend to apply localized patches or complete rewrites when repairing code, and how their behavior diverges from that of human developers. To this end, the authors construct an evaluation benchmark based on Codeforces and present the first quantitative analysis measuring the deviation of LLM repair behaviors from human-generated patches. Furthermore, they conduct automated comparative evaluations across GPT-series models using text similarity algorithms. The findings reveal that LLMs inherently violate the principle of minimal modification, exhibiting a strong preference for reimplementation over targeted patching. Notably, the study demonstrates that generating solutions from scratch outperforms repairing erroneous code. These insights provide critical implications for the design of AI-assisted debugging tools, highlighting fundamental discrepancies between machine-driven and human-driven code repair strategies.

0 citationsRead paper

On the Triangle Inequality for the Jaccard Distance in Arbitrary Lattices

Aug 18, 2026

本文解决了在任意格上推广Jaccard距离的问题，通过证明当估值满足特定条件时Jaccard距离满足三角不等式，并探讨了其在量子信息理论等领域的应用。

0 citationsRead paper

Foundations of MT-PDCL: Measure-Theoretic Probabilistic Definite Clause Logic

Aug 13, 2026

Traditional probabilistic logic programming is constrained by discrete propositionalization, hindering precise reasoning over continuous domains. This work proposes the MT-PDCL framework, which introduces measure theory into probabilistic logic programming for the first time. By leveraging standard Borel σ-algebras and Lebesgue integration, MT-PDCL defines logical variables directly over continuous measurable spaces and establishes a semantics based on continuous probability distributions along with a continuous immediate consequence operator. The approach preserves a purely declarative syntax while enabling exact, algebraic, and structurally differentiable probabilistic inference over continuous domains. Centered on continuous integration rather than discrete combinatorial enumeration, the framework effectively circumvents the limitations of finite domains and supports efficient, exact inference with continuous priors and observations.

0 citationsRead paper

The DIME Architecture: A Unified Operational Algorithm for Neural Representation, Dynamics, Control and Integration

Mar 07, 2026

Current neuroscience lacks a unified computational architecture that integrates core cognitive functions such as perception, memory, prediction, value assessment, and consciousness. This work proposes DIME (Detect-Integrate-Mark-Execute), an operational neural architecture grounded in a four-stage cyclic mechanism that unifies neural representation, dynamic evolution, control, and multiscale integration through the interplay of memory traces, execution trajectories, a marking system, and hyper-memory traces. For the first time, DIME incorporates perception, memory, valuation, and conscious access within a single computational framework, leveraging abstract modeling, recursive processing, neuromodulation, and large-scale network integration. The architecture aligns with empirical findings including hippocampal indexing, cortical replay, and the global workspace theory, while offering a novel, interpretable, and scalable template for next-generation cognitive architectures in artificial intelligence and robotics.

0 citationsRead paper

LoopDB: A Loop Closure Dataset for Large Scale Simultaneous Localization and Mapping

Jun 07, 2025

Existing loop closure detection benchmarks suffer from limited scene diversity, insufficient scale, and sparse ground-truth pose annotations, hindering comprehensive evaluation of SLAM algorithms. To address this, we introduce LoopDB—the first open-source, multi-scene, high-resolution image-sequence benchmark specifically designed for SLAM loop closure detection. LoopDB encompasses six distinct environments (e.g., park, indoor, parking lot, and object close-ups), comprising over 1,000 images organized into 5-frame continuous sequences, with dense, frame-wise relative pose ground truth derived from high-precision camera capture and geometric calibration. The dataset is compatible with mainstream deep learning architectures, including CNNs and Vision Transformers. LoopDB significantly enhances the evaluability of generalization and robustness of loop closure methods under complex, realistic conditions. It is publicly released and has already been adopted by multiple SLAM research groups.

0 citationsRead paper