Zhijiang Li

Hidden-state probes are widely used to study safety-related behavior in language and multimodal models, yet their results can change with the data sources, behavior labels, extracted layer, and probe family. We introduce TriSAIL, an evaluation protocol that records these choices and tests how they affect the resulting claim. TriSAIL uses three response-derived states—benign, non-refusal jailbreak, and borderline/refusal—and assigns train, validation, and test data by source. We extract hidden states from the final input position in the first generation forward pass, before a generated token is returned. The operating layer is selected using validation data only. We evaluate the protocol on five text-only LLMs and six MLLMs with logistic, kNN, and SVM probes. The LLM results vary substantially across held-out attack families, probe families, and source assignments. The MLLM probes achieve high label separability, while source-ID decoding, prediction–source association, and residualization tests show strong source alignment in the same representations. Because the main MLLM matrix couples sources and labels, these results describe source-sensitive separability. Source-transfer conclusions use source-held-out columns; mixed accuracy is a descriptive overall summary. TriSAIL reports these results with probe sensitivity and source diagnostics so that each score is interpreted under the conditions that produced it.

Class imbalance has long posed a significant challenge to semantic segmentation in remote sensing imagery. In semi-supervised settings, existing studies attempt to alleviate this issue through data augmentation, loss function refinement, and optimized training strategies. However, these methods generally rely on an idealized assumption that labeled and unlabeled data share the same class distribution. In practice, unlabeled data typically far exceeds labeled data in scale, making this assumption difficult to satisfy. Such distributional inconsistency not only weakens the effectiveness of semi-supervised learning but also amplifies the inherent class imbalance in remote sensing segmentation. To address these key challenges, we propose a Cross-Prototype Bidirectional Constraint framework (CPBCF) for semi-supervised semantic segmentation. The framework aligns unlabeled prototypes with labeled feature representations to enforce class consistency among unlabeled prototypes, preventing them from being overwhelmed by majority-class features in the early training stages and reducing pseudo-label bias. In addition, by imposing class-level constraints between unlabeled features and labeled prototypes, CPBCF enables the learning of minority-class features to be reinforced by information from unlabeled prototypes, thereby improving the model’s ability to recognize rare categories. To further suppress majority-class dominance, a dynamic class-weight adjustment mechanism is introduced, which selectively applies prototype constraints to minority classes. Through this prototype-level bidirectional balancing strategy, CPBCF enhances the model’s capacity to learn robust minority-class representations under the guidance of unlabeled data, effectively mitigating performance degradation caused by class imbalance in semi-supervised scenarios. Extensive experiments on the ISPRS Potsdam and Vaihingen datasets show that our method consistently surpasses state-of-the-art approaches, achieving notable improvements across all benchmarks.

Cannabichromenic acid synthase (CBCAS) catalyzes a pivotal step in the biosynthesis of cannabichromene (CBC), and its activity level directly affects the production of CBC. In this study, a combined approach involving molecular docking and site-directed mutagenesis was employed to engineer the enzyme, with the goal of generating CBCAS variants that exhibit increased catalytic activity. The CBCAS gene was cloned from a cannabis cultivar with a high CBC content. The three-dimensional structure of the CBCAS protein was predicted using AlphaFold2. Molecular docking between the enzyme and its substrate, cannabigerolic acid (CBGA), was performed using Discovery Studio software. Mutation sites predicted to increase substrate binding affinity were identified, and the resulting recombinant plasmids were constructed and transformed into Pichia pastoris (P. pastoris) to generate engineered strains with increased CBCAS activity. Molecular docking analysis revealed that the CBCAS^Q106Y+E534R and CBCAS^Q124L+E534K double mutants exhibited increased noncovalent interactions and increased hydrophobic contact with the substrate, which likely accounted for their strengthened binding affinity for CBGA. Enzyme activity analysis revealed that compared with those of the wild type (CBCAS^WT), the enzymatic activities of CBCAS^Q106Y+E534R and CBCAS^Q124L+E534K were both significantly elevated (P < 0.05). These mutants catalyzed the conversion of CBGA to yield 93.3 ng/mL and 91.4 ng/mL CBC, representing increases of 37.5% and 34.7%, respectively, relative to CBCAS^WT. Heterologous expression of cannabis CBCAS was successfully achieved in P. pastoris, and mutant enzymes with increased catalytic activity were obtained, laying a foundation for the industrial production and application of recombinant CBCAS.

Remote sensing images are susceptible to atmospheric scattering, imaging conditions, and post-processing strategies during actual acquisition, resulting in issues such as low contrast, insufficient color saturation, and overall poor visual quality. These problems significantly degrade the color quality and expressiveness of the imagery. To address these issues, a prior-guided conditional diffusion enhancement framework (PGCDE) is proposed in this paper. First, an unconditional diffusion model is built upon large-scale natural images to extract stable color priors. Then, these prior features are dynamically injected into the conditional enhancement network through an adaptive hierarchical feature fusion (AHFF) module, with a multi-domain joint loss introduced during training to constrain structural consistency. Finally, at the inference stage, a luminance-decoupled multi-scale fusion strategy is employed to recombine the generated low-frequency color tones with the high-frequency textures of the original image. Experiments on the GID-5 and LoveDA datasets demonstrate that the proposed method outperforms existing representative approaches, providing a practical solution for remote sensing image color quality enhancement that balances perceptual improvement with structure preservation.

Most existing multimodal fusion methods adopt a symmetric fusion strategy to optimize feature consistency between modalities. However, this strategy ignores modality specific discrepancies, and its equal weight distribution mechanism exacerbates the class imbalance problem, resulting in poor classification performance for small-sample categories such as “road”. To address these issues, this study proposes a Category Aware Asymmetric Dynamic Fusion Network (CADFNet), which realizes cross-modal differentiated feature interaction guided by class-level information through the synergy of a Multi-Level Dif ference Pyramid Fusion Module (MDPF) and a Dynamic Sparse Category Attention Module (DSCA). The MDPF module fully mines and enhances cross-modal complementary information by explicitly modeling multi-scale differential features to address the problem of insufficient utilization of modal difference features in existing methods; The DSCA module uses a dynamic sparsity mechanism in conjunction with semantic perception to adaptively assign modal weights to different categories, achieving refined alignment and fusion of cross-modal key areas, thereby effectively alleviating the class imbalance problem. Comprehensive experiments on the PIE-RGB-SAR and WHU-OPT-SAR datasets demonstrate that CADFNet achieves excellent performance. In particular, on the WHU-OPT-SAR dataset, its mean Intersection over Union (mIoU) shows improvements ranging from 1.99% to 15.32% over other state-of-the-art methods.

Optical remote sensing images (RSIs) often suffer from color distortions caused by atmospheric scattering, sensor characteristics, and illumination variations, which degrade visual quality and compromise downstream performance. Non-deep learning methods rely on empirical models and artificial parameters, and are difficult to adapt to complex and changeable remote sensing imaging conditions. Existing CNN-based methods are limited by their local receptive fields, while Transformer-based models incur high computational costs when modeling large-scale RSIs. To address these issues, a Mamba-Conv hybrid multiscale fusion network for RSI color correction is proposed in this work. The network leverages an alternating coupling mechanism between convolution and state-space modeling to unify local detail preservation and global illumination consistency, while a multiscale adaptive fusion strategy ensures coherent cross-scale feature alignment. Extensive experiments show that proposed method consistently surpasses current advanced techniques in both quantitative metrics and visual assessments, offering an effective method to alleviate color distortion in remote sensing image and enhance its interpretability and application value.

Aspect-based sentiment triplet extraction (ASTE) task is a burgeoning subtask of aspect-based sentiment analysis (ABSA), which involves extracting aspect terms, opinion expressions, and related sentiment polarities from texts. However, previous pipeline methods for solving the ASTE task are susceptible to error propagation, while end-to-end sequence labeling methods have not fully utilized the given context. Additionally, existing machine reading comprehension (MRC)-based methods struggle to understand the complex grammatical structure and correspondence between aspect terms and opinion expressions, and fail to fully exploit the deep relationship between aspect-opinion pairs and sentiment polarity. To address these challenges, we propose the PromptReader, a multi-round MRC framework incorporating prompt learning. Specifically, we first design two rounds of MRC-based queries, where each round of queries consists of a pair of static and dynamic queries incorporating part-of-speech features and syntactic dependency information, aiming to jointly learn from two opposite perspectives to identify aspect-opinion pairs. Then, a round of queries based on prompt learning is designed, which contains a pair of dynamic queries, aiming to jointly learn under two different degrees of constraints to better predict sentiment polarity. Comprehensive experiments on four widely recognized benchmark datasets demonstrate that PromptReader surpasses the state-of-the-art methods by a significant margin.