Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning

Existing explainability methods for Large Language Models (LLMs) typically treat hidden states as static points in activation space, assuming that correct and incorrect inferences can be separated using representations from an individual layer. However, these activations are saturated with polysemantic features, leading to linear probes learning surface-level lexical patterns rather than underlying reasoning structures. We introduce Truth as a Trajectory (TaT), which models the transformer inference as an unfolded trajectory of iterative refinements, shifting analysis from static activations to layer-wise geometric displacement. By analyzing displacement of representations across layers, TaT uncovers geometric invariants that distinguish valid reasoning from spurious behavior. We evaluate TaT across dense and Mixture-of-Experts (MoE) architectures on benchmarks spanning commonsense reasoning, question answering, and toxicity detection. Without access to the activations themselves and using only changes in activations across layers, we show that TaT effectively mitigates reliance on static lexical confounds, outperforming conventional probing, and establishes trajectory analysis as a complementary perspective on LLM explainability.

Enhancing Monte Carlo Dropout Performance for Uncertainty Quantification

Knowing the uncertainty associated with the output of a deep neural network is of paramount importance in making trustworthy decisions, particularly in high-stakes fields like medical diagnosis and autonomous systems. Monte Carlo Dropout (MCD) is a widely used method for uncertainty quantification, as it can be easily integrated into various deep architectures. However, conventional MCD often struggles with providing well-calibrated uncertainty estimates. To address this, we introduce innovative frameworks that enhances MCD by integrating different search solutions namely Grey Wolf Optimizer (GWO), Bayesian Optimization (BO), and Particle Swarm Optimization (PSO) as well as an uncertainty-aware loss function, thereby improving the reliability of uncertainty quantification. We conduct comprehensive experiments using different backbones, namely DenseNet121, ResNet50, and VGG16, on various datasets, including Cats vs. Dogs, Myocarditis, Wisconsin, and a synthetic dataset (Circles). Our proposed algorithm outperforms the MCD baseline by 2-3% on average in terms of both conventional accuracy and uncertainty accuracy while achieving significantly better calibration. These results highlight the potential of our approach to enhance the trustworthiness of deep learning models in safety-critical applications.

ETAGE: Enhanced Test Time Adaptation with Integrated Entropy and Gradient Norms for Robust Model Performance

Test time adaptation (TTA) equips deep learning models to handle unseen test data that deviates from the training distribution, even when source data is inaccessible. While traditional TTA methods often rely on entropy as a confidence metric, its effectiveness can be limited, particularly in biased scenarios. Extending existing approaches like the Pseudo Label Probability Difference (PLPD), we introduce ETAGE, a refined TTA method that integrates entropy minimization with gradient norms and PLPD, to enhance sample selection and adaptation. Our method prioritizes samples that are less likely to cause instability by combining high entropy with high gradient norms out of adaptation, thus avoiding the overfitting to noise often observed in previous methods. Extensive experiments on CIFAR-10-C and CIFAR-100-C datasets demonstrate that our approach outperforms existing TTA techniques, particularly in challenging and biased scenarios, leading to more robust and consistent model performance across diverse test scenarios. The codebase for ETAGE is available on https://github.com/afsharshamsi/ETAGE.

Bayesian Low-Rank LeArning (Bella): A Practical Approach to Bayesian Neural Networks

Computational complexity of Bayesian learning is impeding its adoption in practical, large-scale tasks. Despite demonstrations of significant merits such as improved robustness and resilience to unseen or out-of-distribution inputs over their non- Bayesian counterparts, their practical use has faded to near insignificance. In this study, we introduce an innovative framework to mitigate the computational burden of Bayesian neural networks (BNNs). Our approach follows the principle of Bayesian techniques based on deep ensembles, but significantly reduces their cost via multiple low-rank perturbations of parameters arising from a pre-trained neural network. Both vanilla version of ensembles as well as more sophisticated schemes such as Bayesian learning with Stein Variational Gradient Descent (SVGD), previously deemed impractical for large models, can be seamlessly implemented within the proposed framework, called Bayesian Low-Rank LeArning (Bella). In a nutshell, i) Bella achieves a dramatic reduction in the number of trainable parameters required to approximate a Bayesian posterior; and ii) it not only maintains, but in some instances, surpasses the performance of conventional Bayesian learning methods and non-Bayesian baselines. Our results with large-scale tasks such as ImageNet, CAMELYON17, DomainNet, VQA with CLIP, LLaVA demonstrate the effectiveness and versatility of Bella in building highly scalable and practical Bayesian deep models for real-world applications.

Improving MC-Dropout Uncertainty Estimates with Calibration Error-based Optimization

Uncertainty quantification of machine learning and deep learning methods plays an important role in enhancing trust to the obtained result. In recent years, a numerous number of uncertainty quantification methods have been introduced. Monte Carlo dropout (MC-Dropout) is one of the most well-known techniques to quantify uncertainty in deep learning methods. In this study, we propose two new loss functions by combining cross entropy with Expected Calibration Error (ECE) and Predictive Entropy (PE). The obtained results clearly show that the new proposed loss functions lead to having a calibrated MC-Dropout method. Our results confirmed the great impact of the new hybrid loss functions for minimising the overlap between the distributions of uncertainty estimates for correct and incorrect predictions without sacrificing the model's overall performance.

Uncertainty-Aware Credit Card Fraud Detection Using Deep Learning

Countless research works of deep neural networks (DNNs) in the task of credit card fraud detection have focused on improving the accuracy of point predictions and mitigating unwanted biases by building different network architectures or learning models. Quantifying uncertainty accompanied by point estimation is essential because it mitigates model unfairness and permits practitioners to develop trustworthy systems which abstain from suboptimal decisions due to low confidence. Explicitly, assessing uncertainties associated with DNNs predictions is critical in real-world card fraud detection settings for characteristic reasons, including (a) fraudsters constantly change their strategies, and accordingly, DNNs encounter observations that are not generated by the same process as the training distribution, (b) owing to the time-consuming process, very few transactions are timely checked by professional experts to update DNNs. Therefore, this study proposes three uncertainty quantification (UQ) techniques named Monte Carlo dropout, ensemble, and ensemble Monte Carlo dropout for card fraud detection applied on transaction data. Moreover, to evaluate the predictive uncertainty estimates, UQ confusion matrix and several performance metrics are utilized. Through experimental results, we show that the ensemble is more effective in capturing uncertainty corresponding to generated predictions. Additionally, we demonstrate that the proposed UQ methods provide extra insight to the point predictions, leading to elevate the fraud prevention process.

An Uncertainty-Aware Deep Learning Framework for Defect Detection in Casting Products

Defects are unavoidable in casting production owing to the complexity of the casting process. While conventional human-visual inspection of casting products is slow and unproductive in mass productions, an automatic and reliable defect detection not just enhances the quality control process but positively improves productivity. However, casting defect detection is a challenging task due to diversity and variation in defects' appearance. Convolutional neural networks (CNNs) have been widely applied in both image classification and defect detection tasks. Howbeit, CNNs with frequentist inference require a massive amount of data to train on and still fall short in reporting beneficial estimates of their predictive uncertainty. Accordingly, leveraging the transfer learning paradigm, we first apply four powerful CNN-based models (VGG16, ResNet50, DenseNet121, and InceptionResNetV2) on a small dataset to extract meaningful features. Extracted features are then processed by various machine learning algorithms to perform the classification task. Simulation results demonstrate that linear support vector machine (SVM) and multi-layer perceptron (MLP) show the finest performance in defect detection of casting images. Secondly, to achieve a reliable classification and to measure epistemic uncertainty, we employ an uncertainty quantification (UQ) technique (ensemble of MLP models) using features extracted from four pre-trained CNNs. UQ confusion matrix and uncertainty accuracy metric are also utilized to evaluate the predictive uncertainty estimates. Comprehensive comparisons reveal that UQ method based on VGG16 outperforms others to fetch uncertainty. We believe an uncertainty-aware automatic defect detection solution will reinforce casting productions quality assurance.

Confidence Aware Neural Networks for Skin Cancer Detection

Deep learning (DL) models have received particular attention in medical imaging due to their promising pattern recognition capabilities. However, Deep Neural Networks (DNNs) require a huge amount of data, and because of the lack of sufficient data in this field, transfer learning can be a great solution. DNNs used for disease diagnosis meticulously concentrate on improving the accuracy of predictions without providing a figure about their confidence of predictions. Knowing how much a DNN model is confident in a computer-aided diagnosis model is necessary for gaining clinicians' confidence and trust in DL-based solutions. To address this issue, this work presents three different methods for quantifying uncertainties for skin cancer detection from images. It also comprehensively evaluates and compares performance of these DNNs using novel uncertainty-related metrics. The obtained results reveal that the predictive uncertainty estimation methods are capable of flagging risky and erroneous predictions with a high uncertainty estimate. We also demonstrate that ensemble approaches are more reliable in capturing uncertainties through inference.

Objective Evaluation of Deep Uncertainty Predictions for COVID-19 Detection

Deep neural networks (DNNs) have been widely applied for detecting COVID-19 in medical images. Existing studies mainly apply transfer learning and other data representation strategies to generate accurate point estimates. The generalization power of these networks is always questionable due to being developed using small datasets and failing to report their predictive confidence. Quantifying uncertainties associated with DNN predictions is a prerequisite for their trusted deployment in medical settings. Here we apply and evaluate three uncertainty quantification techniques for COVID-19 detection using chest X-Ray (CXR) images. The novel concept of uncertainty confusion matrix is proposed and new performance metrics for the objective evaluation of uncertainty estimates are introduced. Through comprehensive experiments, it is shown that networks pertained on CXR images outperform networks pretrained on natural image datasets such as ImageNet. Qualitatively and quantitatively evaluations also reveal that the predictive uncertainty estimates are statistically higher for erroneous predictions than correct predictions. Accordingly, uncertainty quantification methods are capable of flagging risky predictions with high uncertainty estimates. We also observe that ensemble methods more reliably capture uncertainties during the inference.