Layer pruning is a viable technique for compressing large language models while achieving acceleration proportional to the pruning ratio. In this work, we identify that removing any layer induces a magnitude gap in hidden states, and demonstrate that a simple compensation operation leads to superior performance in iterative layer pruning. This key observation motivates us to propose Prune&Comp, a novel, plug-and-play iterative layer pruning scheme that leverages magnitude compensation to mitigate such gaps in a training-free manner. Specifically, we first estimate the magnitude gap of layer removal and then eliminate it by rescaling the remaining weights offline. We further demonstrate the advantages of Prune&Comp in improving the stability of iterative pruning. When integrated with an iterative prune-and-compensate loop, Prune&Comp consistently enhances existing layer pruning metrics. For instance, when 5 layers of LLaMA-3-8B are pruned with the prevalent Taylor+ metric, Prune&Comp reduces PPL from 512.78 to 16.34 and retains 90.57% of the original performance across 9 question-answering tasks, outperforming the baseline by 24.72%.

Offline reinforcement learning (RL) heavily relies on the coverage of pre-collected data over the target policy’s distribution. Existing studies aim to improve data-policy coverage to mitigate distributional shifts, but overlook security risks from insufficient coverage, and the single-step analysis is not consistent with the multi-step decision-making nature of offline RL. To address this, we introduce the sequence-level concentrability coefficient to quantify coverage, and reveal its exponential amplification on the upper bound of estimation errors through theoretical analysis. Building on this, we propose the Collapsing Sequence-Level Data-Policy Coverage (CSDPC) poisoning attack. Considering the continuous nature of offline RL data, we convert state-action pairs into decision units, and extract representative decision patterns that capture multi-step behavior. We identify rare patterns likely to cause insufficient coverage, and poison them to reduce coverage and exacerbate distributional shifts. Experiments show that poisoning just 1% of the dataset can degrade agent performance by 90%. This finding provides new perspectives for analyzing and safeguarding the security of offline RL.

Acupuncture stimulations in somatosensory system can modulate spatiotemporal brain activity and improve cognitive functions of patients with neurological disorders. The correlation between these somatosensory stimulations and dynamical brain responses is still unclear. We proposed a deep learning framework using electroencephalographic activity of stimulated subjects to decode the needling processes of various acupuncture manipulations performed on Zusanli acupoint. Contrastive representation learning integrated with domain adaptation strategy was applied to estimate 3D hand postures and hand joint motion trajectories of acupuncturist with video recordings, by which finite dimensional representations of behavior manifolds for needling operations were inferred. Distinct transition dynamics of behavior manifold were observed for acupuncture with lifting-thrusting and twisting-rotating manipulations. Moreover, latent neural manifolds of acupuncture evoked EEG signals were estimated in low dimensional state space of brain activities with unsupervised manifold learning, which can reliably represent acupuncture stimulations. Furthermore, a nonlinear decoder based on neural networks was designed to transform neural manifolds to behavior manifolds and further predict acupuncture manipulation as well as needling process. Experimental results demonstrated a high performance of the proposed decoding framework for four types of acupuncture manipulations with a precision of 92.42%. The EEG decoder provides an acupuncture-brain interface linking somatosensory stimulations with neural representations, an effective scheme for revealing clinical efficacy of acupuncture treatment.

RNA editing is a co- or post-transcriptional modification through which some cells can make discrete changes to specific nucleotide sequences within an RNA molecule after transcription. Previous studies found that RNA editing may be critically involved in cancer and aging. However, the function of RNA editing in human early embryo development is still unclear. In this study, through analyzing single cell RNA sequencing data, 36.7% RNA editing sites were found to have a have differential editing ratio among early embryo developmental stages, and there was a great reprogramming of RNA editing rates at the 8-cell stage, at which most of the differentially edited RNA editing sites (99.2%) had a decreased RNA editing rate. In addition, RNA editing was more likely to occur on RNA splicing sites during human early embryo development. Furthermore, long non-coding RNA (lncRNA) editing sites were found more likely to be on RNA splicing sites (odds ratio = 2.19, P = 1.37×10-8), while mRNA editing sites were less likely (odds ratio = 0.22, P = 8.38×10-46). Besides, we found that the RNA editing rate on lncRNA had a significantly higher correlation coefficient with the percentage spliced index (PSI) of lncRNA exons (R = 0.75, P = 4.90×10-16), which indicated that RNA editing may regulate lncRNA splicing during human early embryo development. Finally, functional analysis revealed that those RNA editing-regulated lncRNAs were enriched in signal transduction, the regulation of transcript expression, and the transmembrane transport of mitochondrial calcium ion. Overall, our study might provide a new insight into the mechanism of RNA editing on lncRNAs in human developmental biology and common birth defects.

As an innovative strategy, edge computing has been considered a viable option to address the limitations of cloud computing in supporting the Internet-of-Things applications. However, due to the instability of the network and the increase of the attack surfaces, the security in edge-assisted IoT needs to be better guaranteed. In this paper, we propose an intelligent intrusion detection mechanism, FedACNN, which completes the intrusion detection task by assisting the deep learning model CNN through the federated learning mechanism. In order to alleviate the communication delay limit of federal learning, we innovatively integrate the attention mechanism, and the FedACNN can achieve ideal accuracy with a 50% reduction of communication rounds.