To address the challenges of fine-grained force control and unintuitive human–robot interaction in robotic dexterous manipulation, this paper proposes a semantics-driven bilateral force modulation framework. Methodologically, we introduce the first joint modeling of natural language instructions (e.g., “gently grasp the cup”) with bilateral teleoperation force/motion signals via a multimodal Transformer architecture, integrating the SigLIP language encoder, action tokenization, and fused perception of joint position, velocity, and torque. Our key contributions include: (i) the first end-to-end mapping from linguistic intent to force-level control, enabling real-time, interpretable, and bimanual force modulation; and (ii) empirical validation on single-hand cup-stacking and dual-hand sponge-squeezing tasks, where multi-level force instructions are accurately reproduced. SigLIP significantly improves language–force alignment accuracy, demonstrating the efficacy of semantic-guided imitation learning for force control.

This work addresses the high computational cost of exact range queries and the excessive false positive rates of existing approximate methods such as SRC. To this end, we propose c-DAG, a tunable 1D-Tree–enhanced structure featuring densely overlapping nodes with at least three branches per node, enabling efficient approximate range queries. Theoretical analysis demonstrates that c-DAG reduces the average false positive rate by a logarithmic factor. Furthermore, we introduce a general extension framework tailored to empirical query distributions and prove its superiority over baseline approaches in structural privacy preservation. Experimental evaluation on the Gowalla dataset confirms that c-DAG significantly lowers false positive rates while maintaining comparable time and memory overheads, highlighting its potential for applications in searchable encryption and multimedia retrieval.

To address the problem of missed detections in vision-based BEV 3D object detection caused by low appearance contrast between objects and background, this paper proposes a Region-Oriented Attention (ROA) mechanism. ROA is the first approach to incorporate coarse-grained 2D detection priors into BEV feature learning, explicitly guiding the backbone network to focus on potential object regions. It jointly leverages multi-scale features, large-kernel convolutions, and region-adaptive weighting to enhance sensitivity to small objects while expanding receptive field coverage for large objects. Implemented atop the BEVDet/BEVDepth framework, ROA requires no additional annotations or complex architectural modifications. On the nuScenes benchmark, it achieves absolute improvements of +3.2% in mAP and +2.1% in NDS over the respective baselines, significantly outperforming both BEVDet and BEVDepth. These results demonstrate that region-guided attention effectively enhances the robustness and discriminability of BEV representations.

This study addresses the challenges of lemmatization, dialect identification, and language classification for Romansh—a language encompassing five major dialects and the standardized variety Rumantsch Grischun—by proposing the first unified framework covering all its variants. Leveraging a community-built morphological database, the authors develop a variant-aware, lexicon-based lemmatizer that employs rule-based matching to achieve high lemmatization coverage. This approach is further extended to support dialect identification and language boundary detection. Experimental results demonstrate lemmatization coverage of 77–84% on typical texts, dialect identification accuracy of 95% across a corpus of 30,000 documents, and effective discrimination between Romansh and non-Romansh texts.

This study investigates the indirect employment effects of California’s energy transition on Kern County, a region heavily reliant on the petroleum industry. Although fossil fuel sectors directly employ less than 2% of the local workforce, the analysis—using a structural vector autoregression (SVAR) model integrated with input-output networks and high-frequency monthly employment data—reveals that this industry accounts for 11% of local employment fluctuations. Counterfactual simulations indicate that, absent support from global oil markets, Kern County’s current employment level would be 6.4% lower. The findings uncover significant latent interdependencies within regional supply chains, offering empirical evidence to inform targeted local policies aimed at mitigating socioeconomic vulnerabilities in communities exposed to energy transition risks.

This paper addresses the problem of adaptively designing sampling policies to minimize the average Age of Information (AoI) in status-updating systems where channel statistics—particularly the delay distribution—are unknown and transmissions are unreliable. Specifically, it considers a dual-uncertainty setting: (i) reliable but randomly delayed feedback at the receiver, and (ii) error-prone uplink channels with unknown delay distributions at the sensor. To this end, the AoI optimization is first formulated as a stochastic approximation problem under unknown delay distributions. A Robbins–Monro-type online algorithm with momentum is proposed, which converges almost surely to the optimal threshold-based policy. Theoretically, the algorithm achieves an $O(ln K)$ upper bound on cumulative AoI regret over $K$ epochs, matching the $Omega(ln K)$ minimax lower bound—establishing order optimality. Extensive simulations demonstrate that the proposed algorithm significantly outperforms baseline strategies across diverse delay distributions.

Facing declining participation rates in traditional opinion polling and low credibility of social media data, this paper proposes an AI-driven crowdsourced data credibility assessment framework grounded exclusively in social interaction graph structure—not textual content. Methodologically, we develop a graph-analytic model simulating propagation behaviors of honest versus fraudulent participants, leveraging heterogeneous network structural features to detect multiple types of noncompliant engagement. Crucially, by eschewing semantic dependencies, our approach achieves enhanced cross-platform generalizability. Evaluated on multiple real-world social network datasets, the method attains over 90% accuracy in identifying noncompliant behaviors—significantly outperforming baseline models. Results empirically validate that topological structure alone suffices to effectively model participant eligibility and behavioral patterns. This work establishes a scalable, low-intrusion paradigm for high-credibility public opinion collection.

Fine-grained image recognition (FGIR) has long relied on large-scale pre-trained models, hindering deployment in resource-constrained settings and impeding task-specific architecture design. To address this, we propose TGDA—the first fully from-scratch FGIR framework—eliminating pre-training dependencies via fine-grained-aware teacher-guided data augmentation, knowledge distillation, and weakly supervised learning. We further introduce two lightweight, task-specialized architectures: LRNets, optimized for low-resolution inputs, and ViTFS, an efficient Vision Transformer variant tailored for FGIR. Experiments demonstrate that LRNets achieve a 23% accuracy gain while reducing parameters by 20.6×; ViTFS-T attains competitive or superior performance to state-of-the-art pre-trained methods on CUB and Stanford Cars benchmarks, despite requiring orders-of-magnitude less training data. Our work establishes a new paradigm for pre-training-free FGIR, advancing both efficiency and architectural specialization.

This paper addresses the internal validity quantification of weighted estimators—such as OLS, 2SLS, and TWFE—by asking: under what conditions can such estimators be rigorously interpreted as the average treatment effect (ATE) for some latent subpopulation, and what are the absolute or relative upper bounds on the size of that subpopulation? To answer this, the paper introduces the first systematic “internal validity quantification” framework, grounded in the potential outcomes model and linear projection theory. It derives necessary and sufficient conditions for the existence of a well-defined target subpopulation and provides computable upper bounds on its measure—including closed-form expressions and a general-purpose algorithm. Empirical applications reveal that conventional estimators often represent the ATE for less than 10% of the sample, thereby substantially enhancing transparency, interpretability, and reproducibility in causal inference.

This work addresses the limitations of conventional large language model (LLM) training, which relies on static datasets or fixed reward signals and struggles to continuously expand diverse capabilities. The authors propose the AC/DC framework, which introduces an open-ended co-evolutionary mechanism into LLM development for the first time. By dynamically co-evolving tasks and model competencies within a single training run, AC/DC automatically discovers expert models endowed with novel skills. The approach integrates model merging with natural language–based task generation to synthesize data, thereby sustaining evolving populations of both LLMs and tasks without explicit optimization on downstream benchmarks. The resulting expert models exhibit broader capability coverage than substantially larger models, achieve significant reductions in GPU memory consumption, and demonstrate continual performance gains in multi-agent selection settings.