Selected work

Representative Papers

SteelBench: Evaluating Vision-Language Models in Real-World Industrial Environments

Jul 06, 2026

Existing video benchmarks struggle to evaluate the capability of vision-language models in recognizing worker behaviors and reasoning about safety rules under real-world industrial surveillance conditions—such as low illumination, occlusion, and long-range viewing. This work proposes SteelBench, the first multidimensional diagnostic benchmark tailored to authentic steel plant environments. Constructed from 149 hours of surveillance footage, it comprises 1,345 densely annotated video clips curated via temporal deduplication, category balancing, and visibility-aware sampling. The dataset encompasses actions, personal protective equipment (PPE) attributes, spatial context, and explicit safety rules, and introduces a novel annotation provenance auditing mechanism. Experiments reveal that even the best-performing model achieves only 42.6% accuracy on action recognition (versus 84.6% for humans), exhibits safety judgment error rates of 37–58%, and fails to pass more than two diagnostic tests. Moreover, unaudited model-generated labels can inflate reported accuracy by up to 17 percentage points for related models.

0 citationsRead paper

Attention-Gated Convolutional Networks for Scanner-Agnostic Quality Assessment

Apr 16, 2026

This study addresses the challenge of automated quality control in multi-site, multi-vendor structural MRI data corrupted by motion artifacts. The authors propose an end-to-end model that integrates a hierarchical 2D CNN with a multi-head cross-attention mechanism—a novel application of cross-attention in MRI quality assessment. By jointly capturing local features and modeling global dependencies, the model dynamically focuses on motion-related artifacts such as ringing and blurring while suppressing site-specific confounding variations. Cross-domain experiments on the MR-ART and ABIDE datasets demonstrate strong generalization: the model achieves 0.9920 accuracy on seen sites and maintains a robust 0.755 accuracy across 17 unseen, heterogeneous sites without fine-tuning, significantly advancing out-of-distribution generalization in MRI quality evaluation.

0 citationsRead paper

TechING: Towards Real World Technical Image Understanding via VLMs

Jan 26, 2026

Existing vision-language models (VLMs) struggle to interpret hand-drawn technical diagrams—such as flowcharts and block diagrams—in real-world settings, limiting their applicability in specialized domains. To address this challenge, this work proposes LLaMA-VL-TUG, a novel approach that introduces a synthetic data generation strategy tailored specifically for technical imagery and a new self-supervised pretraining objective. By fine-tuning Llama 3.2 11B-Instruct with this framework, the model achieves a 2.14× improvement in ROUGE-L on synthetic data without requiring large-scale real-world hand-drawn datasets. On authentic hand-drawn diagram benchmarks, it yields the fewest compilation errors across seven out of eight diagram types and demonstrates a 6.97× average increase in F1 score, significantly advancing technical diagram understanding and generation capabilities.

0 citationsRead paper

On the complexity of global Roman domination problem in graphs

Jan 14, 2026

This study investigates the computational complexity of the global Roman domination problem, which seeks a minimum-weight labeling function satisfying the Roman domination condition simultaneously on a graph and its complement. Employing techniques from graph-theoretic modeling, complexity reductions, fine-grained structural analysis, and dynamic programming, the work establishes for the first time that global Roman domination is computationally inequivalent to classical Roman domination. It resolves an open question by proving the problem NP-complete on split graphs and further demonstrates NP-completeness on chordal bipartite graphs, planar bipartite graphs with maximum degree five, and circle graphs. In contrast, the paper presents a linear-time algorithm for cographs, highlighting a tractable case within this otherwise computationally hard landscape.

0 citationsRead paper

AdaGradSelect: An adaptive gradient-guided layer selection method for efficient fine-tuning of SLMs

Dec 12, 2025

To address the trade-off between high computational overhead in fine-tuning small language models (SLMs) and the performance degradation of parameter-efficient methods (e.g., LoRA) due to restrictive low-rank factor spaces, this paper proposes a gradient-driven adaptive layer selection fine-tuning method. Our approach innovatively integrates gradient norm evaluation, Dirichlet prior sampling, and an ε-greedy exploration strategy, enabling broad-layer exploration early in training and progressive focusing on critical Transformer layers as training proceeds. Only parameters of selected sparse layers are updated, while all others remain frozen. Experiments demonstrate that, compared to LoRA (rank 256), our method achieves a 3-percentage-point accuracy gain on GSM8K and matches full-parameter fine-tuning performance on MATH. Moreover, it accelerates training by 12% and reduces GPU memory consumption by 35%.

0 citationsRead paper

Recent publications

Latest Papers

SteelBench: Evaluating Vision-Language Models in Real-World Industrial Environments

Jul 06, 2026

Existing video benchmarks struggle to evaluate the capability of vision-language models in recognizing worker behaviors and reasoning about safety rules under real-world industrial surveillance conditions—such as low illumination, occlusion, and long-range viewing. This work proposes SteelBench, the first multidimensional diagnostic benchmark tailored to authentic steel plant environments. Constructed from 149 hours of surveillance footage, it comprises 1,345 densely annotated video clips curated via temporal deduplication, category balancing, and visibility-aware sampling. The dataset encompasses actions, personal protective equipment (PPE) attributes, spatial context, and explicit safety rules, and introduces a novel annotation provenance auditing mechanism. Experiments reveal that even the best-performing model achieves only 42.6% accuracy on action recognition (versus 84.6% for humans), exhibits safety judgment error rates of 37–58%, and fails to pass more than two diagnostic tests. Moreover, unaudited model-generated labels can inflate reported accuracy by up to 17 percentage points for related models.

0 citationsRead paper

Attention-Gated Convolutional Networks for Scanner-Agnostic Quality Assessment

Apr 16, 2026

This study addresses the challenge of automated quality control in multi-site, multi-vendor structural MRI data corrupted by motion artifacts. The authors propose an end-to-end model that integrates a hierarchical 2D CNN with a multi-head cross-attention mechanism—a novel application of cross-attention in MRI quality assessment. By jointly capturing local features and modeling global dependencies, the model dynamically focuses on motion-related artifacts such as ringing and blurring while suppressing site-specific confounding variations. Cross-domain experiments on the MR-ART and ABIDE datasets demonstrate strong generalization: the model achieves 0.9920 accuracy on seen sites and maintains a robust 0.755 accuracy across 17 unseen, heterogeneous sites without fine-tuning, significantly advancing out-of-distribution generalization in MRI quality evaluation.

0 citationsRead paper

TechING: Towards Real World Technical Image Understanding via VLMs

Jan 26, 2026

Existing vision-language models (VLMs) struggle to interpret hand-drawn technical diagrams—such as flowcharts and block diagrams—in real-world settings, limiting their applicability in specialized domains. To address this challenge, this work proposes LLaMA-VL-TUG, a novel approach that introduces a synthetic data generation strategy tailored specifically for technical imagery and a new self-supervised pretraining objective. By fine-tuning Llama 3.2 11B-Instruct with this framework, the model achieves a 2.14× improvement in ROUGE-L on synthetic data without requiring large-scale real-world hand-drawn datasets. On authentic hand-drawn diagram benchmarks, it yields the fewest compilation errors across seven out of eight diagram types and demonstrates a 6.97× average increase in F1 score, significantly advancing technical diagram understanding and generation capabilities.

0 citationsRead paper

On the complexity of global Roman domination problem in graphs

Jan 14, 2026

This study investigates the computational complexity of the global Roman domination problem, which seeks a minimum-weight labeling function satisfying the Roman domination condition simultaneously on a graph and its complement. Employing techniques from graph-theoretic modeling, complexity reductions, fine-grained structural analysis, and dynamic programming, the work establishes for the first time that global Roman domination is computationally inequivalent to classical Roman domination. It resolves an open question by proving the problem NP-complete on split graphs and further demonstrates NP-completeness on chordal bipartite graphs, planar bipartite graphs with maximum degree five, and circle graphs. In contrast, the paper presents a linear-time algorithm for cographs, highlighting a tractable case within this otherwise computationally hard landscape.

0 citationsRead paper

AdaGradSelect: An adaptive gradient-guided layer selection method for efficient fine-tuning of SLMs

Dec 12, 2025

To address the trade-off between high computational overhead in fine-tuning small language models (SLMs) and the performance degradation of parameter-efficient methods (e.g., LoRA) due to restrictive low-rank factor spaces, this paper proposes a gradient-driven adaptive layer selection fine-tuning method. Our approach innovatively integrates gradient norm evaluation, Dirichlet prior sampling, and an ε-greedy exploration strategy, enabling broad-layer exploration early in training and progressive focusing on critical Transformer layers as training proceeds. Only parameters of selected sparse layers are updated, while all others remain frozen. Experiments demonstrate that, compared to LoRA (rank 256), our method achieves a 3-percentage-point accuracy gain on GSM8K and matches full-parameter fine-tuning performance on MATH. Moreover, it accelerates training by 12% and reduces GPU memory consumption by 35%.

0 citationsRead paper