Selected work

Representative Papers

Variation of sentence length across time and genre

Oct 23, 2018Studies in Corpus Linguistics

This study investigates whether mean sentence length in written English has consistently decreased over time (1810–2000), using the Corpus of Historical American English (COHA), and examines its relationship with grammatical change—particularly the decline of the non-finite purposive marker *in order to*—and genre evolution. Methodologically, it employs large-scale corpus querying, diachronic mixed-effects modeling, multivariate correlation analysis, and genre-normalized frequency computation. Results confirm a statistically significant reduction in mean sentence length; crucially, they reveal a robust cross-temporal association between this reduction and the declining use of *in order to*, with systematic genre-specific variation observed in news and fiction. Moving beyond descriptive sentence-length analyses, the study integrates syntactic change, register evolution, and macro-level sentence-length trends within a unified empirical framework. It thus offers both methodological innovation—through genre-aware diachronic modeling—and theoretical advancement for historical syntax.

17 citations2 influentialRead paper

Learning Wall Segmentation in 3D Vessel Trees using Sparse Annotations

Feb 18, 2025

This work addresses the bottleneck of labor-intensive dense manual annotations in 3D carotid vessel wall segmentation. We propose a fully automatic 3D segmentation framework leveraging only sparse centerline annotations. Methodologically, we introduce a novel bifurcation-axis-perpendicular cross-sectional pseudo-label generation strategy, integrated with centerline-driven transverse slice sampling, adversarial 2D U-Net–based initial segmentation, geometry-aware pseudo-label projection, and end-to-end 3D U-Net optimization. This paradigm efficiently transforms easily obtainable sparse 2D centerline labels into high-fidelity 3D vessel wall segmentations, notably improving accuracy at bifurcations. Evaluated on real carotid imaging data, our method enables precise quantification of 3D biomarkers—including plaque volume—and delivers a clinically deployable, automated solution for carotid stenosis assessment.

1 citationsRead paper

Enhancing the quality of gauge images captured in haze and smoke scenes through deep learning

Oct 04, 2023Optical Engineering + Applications

This study addresses the challenge of low visibility in instrument images caused by smoke and haze, which severely hinders automated meter reading in infrastructure monitoring and emergency response. To this end, the authors construct the first dataset comprising over 14,000 synthetic instrument images, generated using Unreal Engine to simulate realistic smoke and haze degradation. The transferability of two state-of-the-art dehazing networks—FFA-Net and AECR-Net—is systematically evaluated on this dataset. Experimental results demonstrate that AECR-Net achieves superior performance on the synthetic data, attaining an SSIM of 0.98 and a PSNR of 43 dB, significantly enhancing image clarity and thereby effectively supporting downstream automated meter reading tasks.

1 citationsRead paper

Distinguishing case-mix from context heterogeneity in prognostic regression model synthesis settings

Aug 13, 2026

This study addresses the challenge of disentangling whether coefficient heterogeneity in multicenter prognostic models arises from differences in case-mix or site-specific contextual effects. The authors propose the first diagnostic framework capable of decoupling these sources: leveraging an autoencoder to construct a low-dimensional latent space that emphasizes local prognostic relationships, combined with a tailored loss function, site-specific local regressions, and coefficient surface decomposition. This approach separates heterogeneity into a cross-site reference and site-specific deviations, which are then mapped onto the outcome scale to generate both observation-level and site-level summaries. In a two-center COPD trial, the dominant source of slope heterogeneity in key latent variables was attributed to context, while observation-level variance was primarily driven by case-mix; site-level summaries highlighted contextual differences. Negative-control permutation tests confirmed the authenticity of the contextual signal, offering empirical guidance for choosing between unified or localized modeling strategies.

0 citationsRead paper

Attention from Action, for Action: Emergent Visual Bottlenecks for Policy Learning

Aug 13, 2026

This work addresses the limitations of existing visual policy learning methods, which rely on fixed or manually annotated regions of interest and thus struggle to adapt to dynamically shifting visual attention during tasks, resulting in poor data efficiency and robustness. The authors propose Seeker, a self-supervised approach that learns task-aware attention without external annotations by leveraging action signals alone. Built upon frozen DINOv2 features, Seeker iteratively generates dynamic regions of interest conditioned on both state and task, which guide image cropping, background augmentation, and point cloud filtering. Evaluated in both simulated and real-world robotic settings, the method substantially improves performance: task success rates on physical robots increase from 48.3% to 76.7%, and under lighting and background perturbations, from 20.0% to 60.0%.

0 citationsRead paper

Recent publications

Latest Papers

Distinguishing case-mix from context heterogeneity in prognostic regression model synthesis settings

Aug 13, 2026

This study addresses the challenge of disentangling whether coefficient heterogeneity in multicenter prognostic models arises from differences in case-mix or site-specific contextual effects. The authors propose the first diagnostic framework capable of decoupling these sources: leveraging an autoencoder to construct a low-dimensional latent space that emphasizes local prognostic relationships, combined with a tailored loss function, site-specific local regressions, and coefficient surface decomposition. This approach separates heterogeneity into a cross-site reference and site-specific deviations, which are then mapped onto the outcome scale to generate both observation-level and site-level summaries. In a two-center COPD trial, the dominant source of slope heterogeneity in key latent variables was attributed to context, while observation-level variance was primarily driven by case-mix; site-level summaries highlighted contextual differences. Negative-control permutation tests confirmed the authenticity of the contextual signal, offering empirical guidance for choosing between unified or localized modeling strategies.

0 citationsRead paper

Attention from Action, for Action: Emergent Visual Bottlenecks for Policy Learning

Aug 13, 2026

This work addresses the limitations of existing visual policy learning methods, which rely on fixed or manually annotated regions of interest and thus struggle to adapt to dynamically shifting visual attention during tasks, resulting in poor data efficiency and robustness. The authors propose Seeker, a self-supervised approach that learns task-aware attention without external annotations by leveraging action signals alone. Built upon frozen DINOv2 features, Seeker iteratively generates dynamic regions of interest conditioned on both state and task, which guide image cropping, background augmentation, and point cloud filtering. Evaluated in both simulated and real-world robotic settings, the method substantially improves performance: task success rates on physical robots increase from 48.3% to 76.7%, and under lighting and background perturbations, from 20.0% to 60.0%.

0 citationsRead paper

The Role of Variability in Human-Machine Interaction Experience

Aug 11, 2026

This study addresses a key limitation in traditional human-robot shared control: the neglect of natural human movement variability, which often degrades interaction quality and task performance. To overcome this, the authors propose a novel shared optimal controller that explicitly models and incorporates human behavioral variability into the control policy design. The approach is systematically evaluated through haptic interaction experiments comparing three distinct control modes. Results demonstrate that the control mode preserving natural human variability significantly enhances users’ perceived interaction quality while maintaining high task performance. These findings underscore the method’s innovative contribution in effectively balancing usability and efficiency in shared autonomy systems.

0 citationsRead paper

TAF-MED: Multi-Turn Safety Refusal Collapse in LLMs Under Declared Self-Treatment Intent

Aug 10, 2026

This study addresses a critical gap in current evaluations of large language models (LLMs): the sustained adherence to medication safety boundaries in multi-turn dialogues after users explicitly express intent for self-treatment. The authors introduce TAF-MED, a physician-validated benchmark comprising 500 multi-turn scenarios, and evaluate eight LLMs through 4,000 dialogues. Their analysis reveals, for the first time, a “safe refusal collapse” phenomenon—61.4% of initially safe dialogues subsequently produce unsafe responses, with 71.6% of all dialogues containing at least one unsafe reply and model-specific collapse rates ranging from 24.4% to 96.2%. The work argues that safety assessments must consider full dialogue trajectories rather than isolated turns and releases the high-quality TAF-MED benchmark. Using a hybrid approach of automated scoring and dual-physician annotation, the study demonstrates high reliability of automatic evaluation (Cohen’s κ = 0.895, agreement rate 94.3%).

0 citationsRead paper

Circuit-Based Program Verification: Sequential Circuits as an Intermediate Representation for Verifying C Programs

Aug 07, 2026

This work addresses the longstanding divide between software and hardware verification, which has been hindered by the absence of a common intermediate representation that would enable direct application of efficient hardware model checking techniques to C programs. To bridge this gap, the paper introduces the Circuit-based Program Verification (CPV) framework, which systematically compiles C programs into sequential circuits, unifying control-flow and data-flow semantics within a single formal model. CPV integrates established hardware model checking algorithms—including Bounded Model Checking (BMC), k-induction, and IC3/PDR—to support both reachability safety and termination verification. Moreover, it automatically translates counterexamples back into human-readable software evidence. Evaluated on a benchmark suite of over 16,000 verification tasks, CPV matches the performance of leading software verifiers and successfully solves instances beyond the reach of existing tools, demonstrating significant complementary strengths.

0 citationsRead paper