Selected work

Representative Papers

Mixed Gaussian Projections for Two-Sample Testing of Functional Data

Aug 06, 2026

This study addresses the limited power of existing two-sample tests for functional data in finite-sample settings by proposing a hybrid Gaussian random projection method that integrates Haar and Fourier Gaussian components to simultaneously capture local discontinuities and global oscillatory differences. The approach innovatively incorporates a label-invariant, data-adaptive covariance operator, enhancing detection sensitivity while preserving the validity of permutation-based inference. Theoretical analysis establishes the consistency of the proposed test, and empirical evaluations demonstrate its strong specificity and robustness in simulations. When applied to the ECG5000 dataset, the method effectively identifies class distinctions primarily driven by local features and significantly outperforms competing approaches, particularly in scenarios involving changes in covariance structure.

0 citationsRead paper

Meshless Domain Randomization via Explicit Parameter Perturbation of 3D Gaussian Splatting

Jul 24, 2026

Traditional grid-based domain randomization struggles to effectively handle the intricate textures and geometries of complex organisms such as insects, limiting sim-to-real transfer performance. This work proposes the first mesh-free domain randomization framework, which performs explicit perturbations directly in the parameter space of 3D Gaussian splatting. Photometric variations are achieved by modulating spherical harmonics coefficients, original appearances are replaced with 3D procedural noise, and randomized background synthesis is integrated. By eliminating reliance on mesh models, the method provides an efficient data augmentation strategy for complex geometries, substantially narrowing the gap between simulated and real domains and yielding training data with enhanced robustness and superior generalization capabilities.

0 citationsRead paper

BioTIER: A Refusal Benchmark for Targeted Biological Risk Mitigation

Jul 15, 2026

This work addresses critical safety challenges faced by large language models in the biological domain, where models either risk disclosing hazardous information or overly restrict legitimate scientific inquiries due to a lack of precise risk identification and differentiated control mechanisms. To resolve this, the authors propose BioTIER—the first risk-tiered biosafety evaluation benchmark—which categorizes biological content into three distinct classes: catastrophe-averting, dual-use research of concern, and general biology. Leveraging an expert-curated dataset of 542 metadata-annotated prompts, BioTIER enables accurate discrimination between high-risk information and beneficial scientific knowledge. The benchmark facilitates targeted refusal strategies that effectively block a minimal set of catastrophic content while preserving open access to the vast majority of research-relevant knowledge, thereby significantly enhancing both model safety and scientific utility.

0 citationsRead paper

Entropy in Semantic Memory Navigation in Blind and Sighted Individuals: The Effect of Visual Experience

Jul 13, 2026

This study investigates how visual experience shapes the organization and retrieval of semantic memory by comparing congenitally blind individuals with sighted controls in their representation of concrete and abstract concepts. Employing an attribute-generation task, the research introduces— for the first time in cross-sensory semantic memory studies—a semantic entropy metric derived from embedding-based models, complemented by generalized linear mixed-effects modeling for data analysis. Results reveal that sighted participants exhibit higher semantic entropy for abstract concepts, whereas blind individuals show elevated entropy specifically for visually salient concrete concepts (e.g., “penguin”), indicating that visual experience critically influences the dynamic navigation of semantic structure. This work provides novel methodological and empirical insights into the role of sensory experience in shaping semantic representations.

0 citationsRead paper

Comparing Semantic Navigation in Humans and Large Language Models using Natural Language Processing

Jul 13, 2026

This study investigates differences in semantic memory retrieval strategies between humans and large language models (LLMs), specifically examining whether LLMs can replicate the distinctive semantic search patterns observed in human cognition. Using verbal fluency tasks and natural language processing–based semantic trajectory analysis, the research systematically quantifies semantic search dynamics across three complementary dimensions—entropy, step size between adjacent responses, and distance from the semantic centroid—for 82 human participants and three LLMs under varying temperature settings. The findings reveal that humans exhibit higher entropy, larger semantic step sizes, and broader spatial distribution, reflecting a more exploratory search behavior. In contrast, even with temperature tuning, current LLMs fail to simultaneously match human performance across all three dimensions, highlighting fundamental limitations in their ability to emulate human-like semantic navigation mechanisms.

0 citationsRead paper

Recent publications

Latest Papers

Mixed Gaussian Projections for Two-Sample Testing of Functional Data

Aug 06, 2026

This study addresses the limited power of existing two-sample tests for functional data in finite-sample settings by proposing a hybrid Gaussian random projection method that integrates Haar and Fourier Gaussian components to simultaneously capture local discontinuities and global oscillatory differences. The approach innovatively incorporates a label-invariant, data-adaptive covariance operator, enhancing detection sensitivity while preserving the validity of permutation-based inference. Theoretical analysis establishes the consistency of the proposed test, and empirical evaluations demonstrate its strong specificity and robustness in simulations. When applied to the ECG5000 dataset, the method effectively identifies class distinctions primarily driven by local features and significantly outperforms competing approaches, particularly in scenarios involving changes in covariance structure.

0 citationsRead paper

Meshless Domain Randomization via Explicit Parameter Perturbation of 3D Gaussian Splatting

Jul 24, 2026

Traditional grid-based domain randomization struggles to effectively handle the intricate textures and geometries of complex organisms such as insects, limiting sim-to-real transfer performance. This work proposes the first mesh-free domain randomization framework, which performs explicit perturbations directly in the parameter space of 3D Gaussian splatting. Photometric variations are achieved by modulating spherical harmonics coefficients, original appearances are replaced with 3D procedural noise, and randomized background synthesis is integrated. By eliminating reliance on mesh models, the method provides an efficient data augmentation strategy for complex geometries, substantially narrowing the gap between simulated and real domains and yielding training data with enhanced robustness and superior generalization capabilities.

0 citationsRead paper

BioTIER: A Refusal Benchmark for Targeted Biological Risk Mitigation

Jul 15, 2026

This work addresses critical safety challenges faced by large language models in the biological domain, where models either risk disclosing hazardous information or overly restrict legitimate scientific inquiries due to a lack of precise risk identification and differentiated control mechanisms. To resolve this, the authors propose BioTIER—the first risk-tiered biosafety evaluation benchmark—which categorizes biological content into three distinct classes: catastrophe-averting, dual-use research of concern, and general biology. Leveraging an expert-curated dataset of 542 metadata-annotated prompts, BioTIER enables accurate discrimination between high-risk information and beneficial scientific knowledge. The benchmark facilitates targeted refusal strategies that effectively block a minimal set of catastrophic content while preserving open access to the vast majority of research-relevant knowledge, thereby significantly enhancing both model safety and scientific utility.

0 citationsRead paper

Entropy in Semantic Memory Navigation in Blind and Sighted Individuals: The Effect of Visual Experience

Jul 13, 2026

This study investigates how visual experience shapes the organization and retrieval of semantic memory by comparing congenitally blind individuals with sighted controls in their representation of concrete and abstract concepts. Employing an attribute-generation task, the research introduces— for the first time in cross-sensory semantic memory studies—a semantic entropy metric derived from embedding-based models, complemented by generalized linear mixed-effects modeling for data analysis. Results reveal that sighted participants exhibit higher semantic entropy for abstract concepts, whereas blind individuals show elevated entropy specifically for visually salient concrete concepts (e.g., “penguin”), indicating that visual experience critically influences the dynamic navigation of semantic structure. This work provides novel methodological and empirical insights into the role of sensory experience in shaping semantic representations.

0 citationsRead paper

Comparing Semantic Navigation in Humans and Large Language Models using Natural Language Processing

Jul 13, 2026

This study investigates differences in semantic memory retrieval strategies between humans and large language models (LLMs), specifically examining whether LLMs can replicate the distinctive semantic search patterns observed in human cognition. Using verbal fluency tasks and natural language processing–based semantic trajectory analysis, the research systematically quantifies semantic search dynamics across three complementary dimensions—entropy, step size between adjacent responses, and distance from the semantic centroid—for 82 human participants and three LLMs under varying temperature settings. The findings reveal that humans exhibit higher entropy, larger semantic step sizes, and broader spatial distribution, reflecting a more exploratory search behavior. In contrast, even with temperature tuning, current LLMs fail to simultaneously match human performance across all three dimensions, highlighting fundamental limitations in their ability to emulate human-like semantic navigation mechanisms.

0 citationsRead paper