While critical for alignment, Supervised Fine-Tuning (SFT) incurs the risk of catastrophic forgetting, yet the layer-wise emergence of instruction-following capabilities remains elusive.We investigate this mechanism via a comprehensive analysis utilizing informationtheoretic, geometric, and optimization metrics across model scales (1B-32B).Our experiments reveal a distinct depth-dependent pattern: middle layers (20%-80%) are stable, whereas final layers exhibit high sensitivity.Leveraging this insight, we propose Mid-Block Efficient Tuning, which selectively updates these critical intermediate layers.Empirically, our method outperforms standard LoRA up to 10.2% on GSM8K (OLMo2-7B) with reduced parameter overhead, demonstrating that effective alignment is architecturally localized rather than distributed.The code is publicly available at https://github.com/lshowway/base.

The brain patterns of emotional perception remain a pivotal research domain in affective neuroscience. Modeling the brain as a complex network has become a crucial approach to understanding its functions. However, traditional brain network research based on graph theory primarily focuses on dyadic interactions between brain regions, which cannot effectively characterize the information exchange process among multiple brain regions during emotional cognitive processes. To address these limitations, we shift our perspective from graph theory to the topological data analysis (TDA) of minimal 1-cycles. The 1 cycles or loops within a network represent the fundamental high order interactions in complex networks and serve as essential pathways for information transmission and integration among the distributed networks of brain regions. By focusing on cycle structures in affective brain networks, we propose a novel SMA-EL method based on the collaborative optimization of the Minimal 1-Cycle with Simplicial Maps Annotation (SMA-M1C) method and linear programming, which balances computational efficiency and method performance to reconstruct the optimal cycles in the brain network. This method is applied to the analysis of emotional brain networks in response to positive and negative emotions induced by naturalistic viewing. Comprehensive experiments demonstrate that the 1-cycle structures of the brain's functional patterns exhibit differences at both individual and group levels, aligning with prior research. Furthermore, the 1 cycles we proposed can serve as a biological marker for emotion recognition. These findings may provide new insights into the organization patterns of functional brain networks under diverse emotional states.

Image segmentation is a fundamental task in computer vision and computational social systems, with semantic segmentation aiming to assign each pixel to a corresponding label. Due to the inherent richness of categories and contextual information in images, image segmentation remains a challenging problem. Currently, semantic segmentation models based on the segment anything model have demonstrated promising results. However, they continue to encounter challenges related to training strategies and prompt information generation. To address these issues, we propose a straightforward and efficient design method for semantic segmentation based on a prompt-free model, named SAM-SS. First, we introduce the class prompt encoder, which generates category prompts for the mask decoder to extract category-specific semantic information. Second, we incorporate the deep fusion module to bridge the semantic gap for achieving robust representation. Additionally, we observe that fine-tuning the image encoder via low-rank adaptation often leads to suboptimal convergence. To mitigate this, we propose a learning rate modulation strategy to stabilize training and boost model performance. Finally, we validate our model’s performance on three publicly available datasets. Specifically, on the Cityscapes validation set for natural images, our model achieves a mean intersection over union (mIoU) of 85.28%. Moreover, our model demonstrates strong adaptability to remote sensing imagery, achieving mIoU scores of 54.46% on LoveDA and 80.8% on the ISPRS Potsdam validation sets. These results underscore its potential utility across diverse applications.

With the widespread application of large language models (LLM), concerns about the privacy leakage of model training data have increasingly become a focus. Membership Inference Attacks (MIAs) have emerged as a critical tool for evaluating the privacy risks associated with these models. Although existing attack methods, such as LOSS, Reference-based, min-k, and zlib, perform well in certain scenarios, their effectiveness on large pre-trained language models often approaches random guessing, particularly in the context of large-scale datasets and single-epoch training. To address this issue, this paper proposes a novel ensemble attack method that integrates several existing MIAs techniques (LOSS, Reference-based, min-k, zlib) into an XGBoost-based model to enhance overall attack performance (EM-MIAs). Experimental results demonstrate that the ensemble model significantly improves both AUC-ROC and accuracy compared to individual attack methods across various large language models and datasets. This indicates that by combining the strengths of different methods, we can more effectively identify members of the model’s training data, thereby providing a more robust tool for evaluating the privacy risks of LLM. This study offers new directions for further research in the field of LLM privacy protection and underscores the necessity of developing more powerful privacy auditing methods.