representation analysis

This work addresses the critical challenge of evaluating model generalization in high-stakes scenarios with scarce labels, where existing methods lack reliable, label-free metrics for pre-deployment model selection and post-deployment performance monitoring. To bridge this gap, the study introduces, for the first time, the internal causal circuit mechanisms of Vision Transformers into generalization assessment, proposing two novel unsupervised metrics: Dependency Depth Bias and Circuit Shift Score. The former quantifies depth-wise biases in representational dependency structures, while the latter measures changes in circuit stability under distribution shifts. Extensive experiments across diverse tasks demonstrate that these metrics achieve substantially higher correlations with true generalization performance—improving by 13.4% and 34.1% on average over current approaches—thereby significantly enhancing the reliability of generalization prediction without requiring ground-truth labels.

This work addresses the performance degradation of pathological vision-language models (VLMs) under distribution shifts in clinical deployment, a challenge exacerbated by the absence of effective label-free degradation detection mechanisms. To this end, the authors propose a unified monitoring framework that jointly leverages input-level data shift detection and output-level prediction confidence analysis to identify performance deterioration without requiring ground-truth labels. The approach innovatively integrates multi-source unsupervised shift indicators with dynamic confidence metrics within DomainSAT, a lightweight visualization tool. Extensive experiments on large-scale histopathological tumor classification datasets demonstrate that the proposed framework reliably and interpretably detects VLM performance degradation under distributional shifts, thereby significantly enhancing the robustness and trustworthiness of clinical deployment.

This work addresses the challenge of training failures in large language models, which often persist for thousands of optimization steps before manifesting as obvious loss divergence, leading to substantial computational waste. The authors propose a mechanism-aware, proactive monitoring approach that deploys internal detectors at the earliest points where failure signatures become measurable. Specifically, they introduce diagnostic signals grounded in the functional principles of critical modules—such as spectral entropy derived from the bilinear decomposition of QK matrices under low-precision Flash Attention and behavioral metrics of MoE router expert selection. By leveraging these module-specific indicators, the method enables early and accurate identification of diverse failure modes thousands of steps before loss divergence occurs, significantly outperforming conventional detection strategies based solely on loss values or gradient norms, particularly in scenarios involving low-precision attention, excessively high learning rates, or compound faults.

Neural networks frequently exhibit hard-to-diagnose and hard-to-fix anomalous behaviors in production, yet existing maintenance tools are heavily skewed toward the training phase and provide inadequate support for post-deployment diagnosis and root-cause analysis. Method: We adopt a qualitative research approach, conducting in-depth interviews and complementary surveys with 23 practitioners to systematically characterize real-world challenges and tooling gaps in neural network testing, debugging, and maintenance. Contribution/Results: Our study is the first to empirically identify critical shortcomings in current tooling—particularly the neglect of runtime anomaly interpretation, error attribution, and repair validation. We find practitioners urgently require a new maintenance paradigm centered on behavioral observability, causal reasoning, and iterative repair. These findings provide empirical grounding and actionable design principles for building next-generation neural network maintenance infrastructures.

This work addresses the unclear internal mechanisms underlying the failure of deep neural networks to generalize on unseen samples, particularly the lack of effective characterization of shifts in internal decision logic. The authors propose a novel perspective termed “Decision Pattern Shift” (DPS), which for the first time links generalization performance to the stability of internal decision processes. Specifically, they construct channel contribution vectors via GradCAM to represent the decision logic of individual samples and quantify generalization failure by measuring the deviation of these vectors from class-wise average patterns. The proposed framework offers a unified explanation for diverse generalization degradation scenarios, enabling early risk detection and precise defect localization. Experiments demonstrate that DPS exhibits a strong linear correlation with generalization gap (Pearson r > 0.8) and organizes various degradation cases into a continuous spectrum, revealing a systematic drift mechanism behind generalization failure.

This work addresses the limitation of existing output-confidence–based fault detection methods, which often fail to capture internal errors in neural networks. The authors propose Self-Detecting Neural Networks (SDNN), a novel framework that introduces the concept of “spectral drift” to reveal that erroneous predictions manifest as pronounced multi-scale spectral instabilities in internal activations. Spectral features are extracted via short-time Fourier transform, wavelet decomposition, and statistical moments, and a lightweight detector is trained using curriculum learning to establish an end-to-end learnable internal monitoring mechanism. Evaluated on CIFAR-10, SDNN achieves an AUROC of 79.0 ± 25.3%, outperforming baseline methods such as MaxSoftmax and Energy Score by 25–30 percentage points.

This study addresses the coupled challenges of class imbalance and class overlap in software defect prediction, which jointly impair model training dynamics and performance. The authors propose the first interaction-aware protocol for analyzing training dynamics under these intertwined data quality issues. By training a fixed multilayer perceptron (MLP) under three conditions—imbalance only, overlap only, and their coupling—the protocol systematically records training trajectories. Integrating effect size analysis, sensitivity analysis, and rule-based classification, it constructs a taxonomy of training dynamic patterns. The work uncovers distinctive neural network behaviors specific to the coupled scenario, offering empirical insights and novel diagnostic tools to enhance the understanding, evaluation, and refinement of defect prediction models.

This study addresses the lack of systematic evaluation of existing test selection metrics under multi-objective settings, distribution shifts, and multimodal data—challenges that hinder practical metric selection. To bridge this gap, the authors construct the first unified benchmark encompassing three testing objectives (fault detection, performance estimation, and retraining guidance), five types of distribution shifts, three data modalities (images, text, and Android packages), and 13 deep learning models. Through a large-scale empirical study involving 1,640 experimental scenarios, they conduct rigorous statistical analyses to comprehensively compare the performance of 15 widely used metrics, elucidate their respective applicability boundaries, and provide reliable guidance and actionable recommendations for test selection in safety-critical systems.

This study investigates the reliability of frozen safety activation monitors—originally trained on base language models—after the models undergo common post-training updates such as quantization, full fine-tuning, or LoRA-based adaptation. Through systematic evaluation across multiple open-source large language models and architectures, the work reveals distinct impacts of these update strategies on monitor performance: quantization (e.g., NF4) has minimal effect, whereas fine-tuning—particularly QLoRA—frequently compromises monitor efficacy, with privacy-related probes being most vulnerable and compliance-based refusal monitors relatively robust. The paper further introduces a method to predict monitor degradation based on pre-deployment model characteristics, enabling efficient prioritization of revalidation efforts and enhancing the safety and efficiency of model deployment.

This work addresses the performance bottlenecks and limited trustworthiness of deep learning models in industrial surface defect detection, which stem from their strong data dependency and poor interpretability. To this end, the authors propose a knowledge-guided loss function that integrates interpretability into the training process without incurring additional inference overhead. The approach employs a two-stage training strategy: first, a primary classification network is trained and used to generate saliency maps as prior knowledge; then, a multi-task learning framework is constructed, where an auxiliary task enforces consistency between the saliency maps of the final model and those of the primary model. Experiments on multiple public defect datasets demonstrate that the proposed method significantly improves baseline models in terms of accuracy and average precision (AP), while producing more focused and human-aligned saliency maps, thereby achieving a synergistic optimization of performance and interpretability.