Large language models (LLMs) have achieved remarkable success in various natural language processing tasks, yet they remain prone to generating factually incorrect outputs—known as "hallucinations". While recent approaches have shown promise for hallucination detection by repeatedly sampling from LLMs and quantifying the semantic inconsistency among the generated responses, they rely on fixed sampling budgets that fail to adapt to query complexity, resulting in computational inefficiency. We propose an Adaptive Bayesian Estimation framework for Semantic Entropy with Guided Semantic Exploration, which dynamically adjusts sampling requirements based on observed uncertainty. Our approach employs a hierarchical Bayesian framework to model the semantic distribution, enabling dynamic control of sampling iterations through variance-based thresholds that terminate generation once sufficient certainty is achieved. We also develop a perturbation-based importance sampling strategy to systematically explore the semantic space. Extensive experiments on four QA datasets demonstrate that our method achieves superior hallucination detection performance with significant efficiency gains. In low-budget scenarios, our approach requires about 50% fewer samples to achieve comparable detection performance to existing methods, while delivers an average AUROC improvement of 12.6% under the same sampling budget.

Instruction Fine-Timing (IFT) is crucial for aligning large language models (LLMs) with human preferences, and selecting a small yet representative subset from massive data significantly facilitates IFT in terms of both efficiency and effectiveness. Nevertheless, existing approaches suffer from two limitations: the use of simple heuristics restricts data diversity, while the singleton data quality evaluation accounts for inconsistent criteria between independent samples. To address the issues, we present TACOS, an innovative method that integrates Open Tagging and Comparative Scoring for IFT data selection. To capture data diversity, we leverage LLMs to assign open-domain tags to human queries, followed by a normalization stage to denoise the open tags and enable efficient clustering. Additionally, we suggest a comparative scoring method that allows the relative quality evaluation of samples within a cluster, avoiding inconsistent criteria seen in singleton-based evaluations. Extensive experiments across diverse datasets and LLM architectures demonstrate that TACOS outperforms existing approaches by a large margin. Notably, it achieves superior instruction-following performance on MT-Bench and ranks 1st among LLaMA2-7B-Based models on AlpacaEval 2.0, illustrating its efficacy for IFT data selection.

Serial section electron microscopy (ssEM) is a pivotal technique for investigating neuronal connections and brain microstructures. However, imperfect sample preparation and image acquisition often lead to degradation, posing challenges for subsequent analysis. While previous deep learning methods, such as the interpolation model using spatially adaptive convolutions, have been proven to outperform conventional approaches, they struggle to recover high-frequency details, resulting in poor perceptual quality and segmentation performance. This study presents a novel approach leveraging diffusion models to restore missing slices of ssEM images. To accommodate the anisotropic characteristic of ssEM images, we enhance the backbone network with asymmetric and symmetric 3D convolutions. Additionally, we propose the Adaptive and Learnable Reconstruction (ALR) module with the First and Last slices Attention Block (FLAB) for effective feature extraction. A Multi-output Joint Strategy (MJS) is utilized for noise estimation, reducing training-testing discrepancies and achieving diffusion correction. Moreover, we also redesign the inference process to optimize the restoration of partially damaged slices, enabling restoration without additional artifact simulation or retraining. Experiment results demonstrate the effectiveness of our approach in generating more realistic slices and its superior performance in downstream tasks, surpassing previous methods.

Many researches have been carried out to explore the ultrastructure of organisms at home and abroad, and different algorithms have been used to reconstruct the images of different samples at the nanoscale. Image “matching” and “fusion” are two significant research fields that directly affect the performance of image Mosaic. As image stitching the first and the last step, if there is no correct image matching and fusion algorithm, it is almost impossible to successful image stitching. Image registration is to map one image to another by looking for a space transformation for two images in a set of image data, so that the points corresponding to the same position in the space in the two images correspond one to one, so as to achieve the purpose of information fusion. In this paper, the feature-based image stitching algorithm was used to transform the analysis of the whole image into the analysis of some features of the image, which greatly reduced the calculation amount. Meanwhile, the feature points were further optimized and selected to reduce the mismatching rate, and the electron microscope images of autistic mice were mosaically splicing. In the registration part, this paper adopts the elastic registration method and makes fine adjustments on this basis to obtain more accurate registration results, which will be helpful for the subsequent segmentation of each organizational structure.