Selected work

Representative Papers

Rules or Character? Scaling Laws for AI Safety Design

Aug 13, 2026

This study investigates how AI systems should dynamically balance rule-based safety mechanisms against behavior-shaping approaches as deployment scale increases, aiming to minimize both expected harm and tail risk. We formalize this trade-off for the first time, introducing a parameter α to represent the allocation of resources between the two strategies. Incorporating factors such as filter degradation, common-mode failures, and baseline behavioral vulnerability, we evaluate the trade-off using comparative statics, a multiplicative Pareto damage model, Monte Carlo simulations, and Conditional Value-at-Risk (CVaR). Our results show that the baseline vulnerability of behavior shaping is the dominant determinant of the optimal strategy. As deployment scale grows, the optimal α* shifts modestly to substantially toward behavior shaping (Δα* = +0.01 to +0.21), reaching offsets up to 0.50 under high vulnerability; moreover, at large scales, the CVaR-optimal and expected-harm-optimal solutions converge.

0 citationsRead paper

Random Window Augmentations for Deep Learning Robustness in CT and Liver Tumor Segmentation

Oct 09, 2025

Direct application of natural-image augmentation methods to CT segmentation disregards the physical meaning of Hounsfield Unit (HU) values, leading to artifacts and poor generalizability. To address this, we propose Random Window-Width Augmentation (RWWA), a CT-specific intensity augmentation method that dynamically samples window width and level based on the empirical HU distribution—thereby preserving anatomical interpretability and physical consistency of HU values. RWWA significantly improves model robustness to low-contrast and multiphase CT images. As the first work to systematically expose the limitations of generic intensity augmentation in CT and introduce a modality-adapted enhancement strategy, RWWA achieves state-of-the-art performance across multiple multi-center liver tumor segmentation benchmarks. Notably, it yields average Dice score improvements of 2.3–4.1% on challenging cases, empirically validating the critical role of physics-aware augmentation in enhancing model generalization.

0 citationsRead paper

Recent publications

Latest Papers

Rules or Character? Scaling Laws for AI Safety Design

Aug 13, 2026

This study investigates how AI systems should dynamically balance rule-based safety mechanisms against behavior-shaping approaches as deployment scale increases, aiming to minimize both expected harm and tail risk. We formalize this trade-off for the first time, introducing a parameter α to represent the allocation of resources between the two strategies. Incorporating factors such as filter degradation, common-mode failures, and baseline behavioral vulnerability, we evaluate the trade-off using comparative statics, a multiplicative Pareto damage model, Monte Carlo simulations, and Conditional Value-at-Risk (CVaR). Our results show that the baseline vulnerability of behavior shaping is the dominant determinant of the optimal strategy. As deployment scale grows, the optimal α* shifts modestly to substantially toward behavior shaping (Δα* = +0.01 to +0.21), reaching offsets up to 0.50 under high vulnerability; moreover, at large scales, the CVaR-optimal and expected-harm-optimal solutions converge.

0 citationsRead paper

Random Window Augmentations for Deep Learning Robustness in CT and Liver Tumor Segmentation

Oct 09, 2025

Direct application of natural-image augmentation methods to CT segmentation disregards the physical meaning of Hounsfield Unit (HU) values, leading to artifacts and poor generalizability. To address this, we propose Random Window-Width Augmentation (RWWA), a CT-specific intensity augmentation method that dynamically samples window width and level based on the empirical HU distribution—thereby preserving anatomical interpretability and physical consistency of HU values. RWWA significantly improves model robustness to low-contrast and multiphase CT images. As the first work to systematically expose the limitations of generic intensity augmentation in CT and introduce a modality-adapted enhancement strategy, RWWA achieves state-of-the-art performance across multiple multi-center liver tumor segmentation benchmarks. Notably, it yields average Dice score improvements of 2.3–4.1% on challenging cases, empirically validating the critical role of physics-aware augmentation in enhancing model generalization.

0 citationsRead paper