As hateful and misleading content continues to proliferate on social media platforms, detecting multimodal harmful content in micro-videos has become increasingly important. While recent advances in this area have shown promise, existing methods often assume complete access to all modalities, including video, audio, and textual descriptions. This assumption fundamentally limits their robustness in real-world settings, where micro-videos are frequently modality- and segment-incomplete due to upload artifacts, platform-level compression, or missing metadata. To bridge this gap, we introduce three real-world benchmarks that systematically capture varying degrees of incompleteness at both the modality and segment levels. We further propose MVKD, a multi-view knowledge distillation framework tailored for robust harmful micro-video detection under incomplete and uncertain conditions. MVKD (1) adaptively distills modality-specific knowledge through confidence-aware guidance, (2) enhances sample- and domain-level representations via contrastive alignment, and (3) integrates complementary cues using a multi-view mixture-of-experts strategy. Extensive experiments show that MVKD consistently outperforms strong baselines in both accuracy and robustness across diverse missingness patterns, and generalizes effectively to previously unseen domains.
Accurate cropland change detection from high-resolution remote sensing images is essential for agricultural resource monitoring, land management, and food security. Existing methods struggle with two primary challenges: (1) severe pseudo-change noise and false alarms, caused by sole reliance on simple spatial or temporal difference enhancement, and (2) inadequate fine-grained cropland change detection, resulting from boundary detail degradation during fixed interpolation decoding. To address these issues, we propose a SpatioTemporal Difference-Enhancement Network (STDENet) with semantic flow alignment for effective cropland change detection from high-resolution remote sensing images. Built on a pyramid vision transformer V2 (PVT-V2) encoder, STDENet integrates three key innovations: a spatiotemporal dual-branch difference enhancement (STDDE) module that jointly captures temporal variations and spatial discrepancies to suppress irrelevant responses from unchanged or spectrally similar regions; an adaptive boundary-aware (ABA) module that dynamically extracts four-directional texture cues to enhance boundary representations; and a semantic flow-aligned decoding (SFAD) module that predicts semantic flow fields to warp high-level difference features and align them with low-level spatial details for fine-grained structure recovery. We evaluate the STDENet on one self-constructed dataset, DYCD, and two public benchmark datasets, CLCD and Hi-CNA. Comparative experiments confirm that STDENet consistently outperforms existing state-of-the-art methods on all three datasets, improving IoU by 3.41 %, 4.90 %, and 1.98 % over the second-best methods on DYCD, CLCD, and Hi-CNA, respectively. By enhancing global structural information through the spatial branch and local change responses through the temporal branch, the spatiotemporal interaction design improves the feature separability of cropland changes. Semantic flow alignment helps restore fine-grained details by producing a more stable decoding process for local contours in cropland change patches. STDENet also demonstrates stronger resistance to interference and better adaptability to phenological variations among crop growth, pre-sowing, and post-harvest periods. Large-scale applications further indicate its generalization ability across heterogeneous cropland structures.
Learning performance modeling has achieved strong predictive performance, but the decision processes of neural models often remain implicit and difficult to analyze. To address this issue, we propose SPARK, a neural-to-symbolic framework for recovering symbolic decision structures from neural predictors of learning performance. SPARK derives performance-related factors from educational interaction data, learns the decision function with a multilayer perceptron as the teacher model, transfers it to a Kolmogorov-Arnold Network (KAN), and then applies structural refinement and symbolic regression to recover compact symbolic expressions. The goal is to convert implicit neural decision behavior into an analyzable symbolic form while preserving its main predictive behavior. Experiments on benchmark educational datasets show that SPARK remains competitive in prediction, preserves most of the teacher model’s decision behavior, and reveals dataset-dependent interaction patterns among student ability, cognition-related factor, and item difficulty.
Artificial intelligence agents are increasingly deployed as professional teammates. However, it remains unclear how they can enhance collaborative performance without hindering human skill development, particularly for novices in high-stakes domains. Cognitive offloading, using external resources (in this case, AI assistants) to reduce one’s own cognitive effort, is central to this tension.We conducted a controlled experiment using a custom platform and the CAIL 2018 dataset. 130 participants were randomly assigned to one of four conditions: a no-AI control, high-offloading AI (which provided direct recommendation), moderate-offloading AI (which provided analytical support), or low-offloading AI (which provided evaluative feedback). The results reveal a fundamental trade-off. The high-offloading condition achieved the highest decision accuracy at 68.8% and the shortest decision time (37.55s). However, it diminished skill development, showing a 7.1% lower accuracy improvement than the control. In contrast, moderate- and low-offloading conditions fostered skill gains of approximately 5.5%. Mediation analysis indicated that this impeded skill development was associated with metacognitive miscalibration (a gap between perceived and actual ability) rather than merely reduced behavioral engagement. These findings extend the “AI Competency Paradox” by revealing a critical offloading threshold below which AI can enhance collaborative performance while supporting skill development, and by identifying metacognitive miscalibration (not reduced effort) as a mechanism underlying skill impairment, suggesting that AI assistants for novices should prioritize adaptive cognitive friction and real-time competence calibration over full automation.
This study examines how the state and the market shape income divisions among paid care workers in China's domestic service economy. We develop the concept of the state-led reproductive boundary to examine how state intervention, through certification, training, work arrangement, and long-term care insurance, redraws the lines along which reproductive labour is valued, financed, and protected. Drawing on a 2021 survey of 919 domestic workers in Shanghai, the first Chinese city to implement a municipality-wide long-term care insurance scheme, we distinguish maternity-infant care, eldercare, and non-nurturant work, and model both monthly and hourly income. We compare a purely cultural account of care valuation with one that incorporates the state. Cultural valuation alone cannot explain the wage divisions within paid care. The division between nurturant and non-nurturant work in China is organised less through race than through human capital, certification, and market recognition. Within nurturant work, the maternity-infant care advantage holds in monthly but not hourly terms, since live-in arrangements obscure the low hourly returns of feminised, time-intensive care. The culturally anticipated eldercare penalty does not materialise. Eldercare workers absorbed into the hourly-priced, state-financed insurance scheme are no longer disadvantaged relative to non-nurturant workers, showing that state intervention can partially insulate a culturally devalued form of care from market discounting.
