Selected work

Representative Papers

JavaVulBench: A Java Vulnerability Benchmark with Realistic Splits, a Unified Multi-Backend Harness, and a Leakage-Aware Evaluation Mode

Jul 02, 2026

This work addresses the critical limitations of existing Java vulnerability detection benchmarks, which inadequately balance realistic scenario partitioning, evaluation consistency, and data leakage prevention. To bridge this gap, we construct a high-quality dataset comprising approximately 30,600 methods spanning 1,740 CVEs, introduce a leakage-aware evaluation protocol alongside a pretraining contamination auditing mechanism, and design five practical data splitting strategies. We further establish the first unified evaluation framework that seamlessly integrates encoder-only models, local generative models, and API-based large language models, enabling single-command, multi-backend cross-model assessment for CodeBERT, GraphCodeBERT, UniXcoder, DeepSeek-Coder, and major LLM APIs. The project releases all data, code, fine-tuned models, and twelve reference detectors, accompanied by a complete evaluation pipeline demonstration.

0 citationsRead paper

Beyond Journals: Rethinking Research Evaluation in Hungarian Computer Science

Jun 09, 2026

This study addresses the misalignment between Hungary’s research evaluation system and global norms in computer science, where excessive reliance on journal publications and neglect of top-tier international conferences distort scholarly incentives. For the first time, it systematically integrates multiple data sources—including iCore, DBLP, MTMT, and MTA-ATT—to quantitatively assess Hungarian researchers’ output, disciplinary distribution, and career trajectories in CORE A* and A conferences through bibliometric analysis, author disambiguation, and temporal topic modeling. Findings reveal that theoretical subfields adopted conference publishing earlier, while many high-performing researchers have emigrated. The study recommends aligning national evaluation criteria with international standards by equating CORE A* conferences with D1 journals and CORE A conferences with Q1 journals, thereby fostering a more globally integrated and incentive-compatible research assessment framework.

0 citationsRead paper

Vision Transformers Need Better Token Interaction

May 22, 2026

This work addresses the issue of semantic diffusion in local representations of Vision Transformers under prolonged training, which degrades performance in dense prediction tasks. To mitigate this over-smoothing effect while preserving global context, the authors propose replacing the standard softmax attention with entmax-1.5 sparse attention, thereby enhancing semantic focus on local tokens. Implemented within the DINOv1 ViT-S/16 architecture, the method maintains the complementary relationship between the [CLS] token and local features. Without compromising linear probing accuracy on ImageNet-1K, it achieves substantial gains in segmentation performance, yielding mIoU scores of 48.78 on PASCAL VOC, 21.97 on ADE20K, and 37.87 on Cityscapes.

0 citationsRead paper

Towards Better Static Code Analysis Reports: Sentence Transformer-based Filtering of Non-Actionable Alerts

Apr 20, 2026

Static code analysis tools often suffer from generating excessive non-actionable warnings, leading to developer alert fatigue and reduced practical utility. This work proposes STAF, a novel approach that, for the first time, leverages Sentence Transformers to classify the actionability of static analysis alerts. By employing sentence embeddings within a binary classification framework, STAF is trained and evaluated on a large-scale dataset of Java projects. Experimental results demonstrate that STAF achieves an F1 score of 89% in within-project settings, outperforming existing methods by at least 11%. In cross-project scenarios, it improves performance by at least 6%, substantially reducing non-actionable alerts and significantly enhancing the overall quality of analysis reports.

0 citationsRead paper

Why Avoid Generative Legal AI Systems? Hallucination, Overreliance, and their Impact on Explainability

Mar 16, 2026

针对生成式法律AI因幻觉与过度依赖削弱可解释性的问题，文章通过分析其技术局限与治理框架，主张审慎部署以保障司法独立与基本权利。

0 citationsRead paper

Recent publications

Latest Papers

JavaVulBench: A Java Vulnerability Benchmark with Realistic Splits, a Unified Multi-Backend Harness, and a Leakage-Aware Evaluation Mode

Jul 02, 2026

This work addresses the critical limitations of existing Java vulnerability detection benchmarks, which inadequately balance realistic scenario partitioning, evaluation consistency, and data leakage prevention. To bridge this gap, we construct a high-quality dataset comprising approximately 30,600 methods spanning 1,740 CVEs, introduce a leakage-aware evaluation protocol alongside a pretraining contamination auditing mechanism, and design five practical data splitting strategies. We further establish the first unified evaluation framework that seamlessly integrates encoder-only models, local generative models, and API-based large language models, enabling single-command, multi-backend cross-model assessment for CodeBERT, GraphCodeBERT, UniXcoder, DeepSeek-Coder, and major LLM APIs. The project releases all data, code, fine-tuned models, and twelve reference detectors, accompanied by a complete evaluation pipeline demonstration.

0 citationsRead paper

Beyond Journals: Rethinking Research Evaluation in Hungarian Computer Science

Jun 09, 2026

This study addresses the misalignment between Hungary’s research evaluation system and global norms in computer science, where excessive reliance on journal publications and neglect of top-tier international conferences distort scholarly incentives. For the first time, it systematically integrates multiple data sources—including iCore, DBLP, MTMT, and MTA-ATT—to quantitatively assess Hungarian researchers’ output, disciplinary distribution, and career trajectories in CORE A* and A conferences through bibliometric analysis, author disambiguation, and temporal topic modeling. Findings reveal that theoretical subfields adopted conference publishing earlier, while many high-performing researchers have emigrated. The study recommends aligning national evaluation criteria with international standards by equating CORE A* conferences with D1 journals and CORE A conferences with Q1 journals, thereby fostering a more globally integrated and incentive-compatible research assessment framework.

0 citationsRead paper

Vision Transformers Need Better Token Interaction

May 22, 2026

This work addresses the issue of semantic diffusion in local representations of Vision Transformers under prolonged training, which degrades performance in dense prediction tasks. To mitigate this over-smoothing effect while preserving global context, the authors propose replacing the standard softmax attention with entmax-1.5 sparse attention, thereby enhancing semantic focus on local tokens. Implemented within the DINOv1 ViT-S/16 architecture, the method maintains the complementary relationship between the [CLS] token and local features. Without compromising linear probing accuracy on ImageNet-1K, it achieves substantial gains in segmentation performance, yielding mIoU scores of 48.78 on PASCAL VOC, 21.97 on ADE20K, and 37.87 on Cityscapes.

0 citationsRead paper

Towards Better Static Code Analysis Reports: Sentence Transformer-based Filtering of Non-Actionable Alerts

Apr 20, 2026

Static code analysis tools often suffer from generating excessive non-actionable warnings, leading to developer alert fatigue and reduced practical utility. This work proposes STAF, a novel approach that, for the first time, leverages Sentence Transformers to classify the actionability of static analysis alerts. By employing sentence embeddings within a binary classification framework, STAF is trained and evaluated on a large-scale dataset of Java projects. Experimental results demonstrate that STAF achieves an F1 score of 89% in within-project settings, outperforming existing methods by at least 11%. In cross-project scenarios, it improves performance by at least 6%, substantially reducing non-actionable alerts and significantly enhancing the overall quality of analysis reports.

0 citationsRead paper

Why Avoid Generative Legal AI Systems? Hallucination, Overreliance, and their Impact on Explainability

Mar 16, 2026

针对生成式法律AI因幻觉与过度依赖削弱可解释性的问题，文章通过分析其技术局限与治理框架，主张审慎部署以保障司法独立与基本权利。

0 citationsRead paper