Shuang Ao

Shuang Ao, Stefan Rueger, Advaith Siddharthan

Proper confidence calibration of deep neural networks is essential for reliable predictions in safety-critical tasks. Miscalibration can lead to model over-confidence and/or under-confidence; i.e., the model's confidence in its prediction can be greater or less than the model's accuracy. Recent studies have highlighted the over-confidence issue by introducing calibration techniques and demonstrated success on various tasks. However, miscalibration through under-confidence has not yet to receive much attention. In this paper, we address the necessity of paying attention to the under-confidence issue. We first introduce a novel metric, a miscalibration score, to identify the overall and class-wise calibration status, including being over or under-confident. Our proposed metric reveals the pitfalls of existing calibration techniques, where they often overly calibrate the model and worsen under-confident predictions. Then we utilize the class-wise miscalibration score as a proxy to design a calibration technique that can tackle both over and under-confidence. We report extensive experiments that show our proposed methods substantially outperforming existing calibration techniques. We also validate our proposed calibration technique on an automatic failure detection task with a risk-coverage curve, reporting that our methods improve failure detection as well as trustworthiness of the model. The code are available at \url{https://github.com/AoShuang92/miscalibration_TS}.

Shuang Ao, Stefan Rueger, Advaith Siddharthan

Failure detection (FD) in AI systems is a crucial safeguard for the deployment for safety-critical tasks. The common evaluation method of FD performance is the Risk-coverage (RC) curve, which reveals the trade-off between the data coverage rate and the performance on accepted data. One common way to quantify the RC curve by calculating the area under the RC curve. However, this metric does not inform on how suited any method is for FD, or what the optimal coverage rate should be. As FD aims to achieve higher performance with fewer data discarded, evaluating with partial coverage excluding the most uncertain samples is more intuitive and meaningful than full coverage. In addition, there is an optimal point in the coverage where the model could achieve ideal performance theoretically. We propose the Excess Area Under the Optimal RC Curve (E-AUoptRC), with the area in coverage from the optimal point to the full coverage. Further, the model performance at this optimal point can represent both model learning ability and calibration. We propose it as the Trust Index (TI), a complementary evaluation metric to the overall model accuracy. We report extensive experiments on three benchmark image datasets with ten variants of transformer and CNN models. Our results show that our proposed methods can better reflect the model trustworthiness than existing evaluation metrics. We further observe that the model with high overall accuracy does not always yield the high TI, which indicates the necessity of the proposed Trust Index as a complementary metric to the model overall accuracy. The code are available at \url{https://github.com/AoShuang92/optimal_risk}.

Shuang Ao, Stefan Rueger, Advaith Siddharthan

Despite the great success of state-of-the-art deep neural networks, several studies have reported models to be over-confident in predictions, indicating miscalibration. Label Smoothing has been proposed as a solution to the over-confidence problem and works by softening hard targets during training, typically by distributing part of the probability mass from a `one-hot' label uniformly to all other labels. However, neither model nor human confidence in a label are likely to be uniformly distributed in this manner, with some labels more likely to be confused than others. In this paper we integrate notions of model confidence and human confidence with label smoothing, respectively \textit{Model Confidence LS} and \textit{Human Confidence LS}, to achieve better model calibration and generalization. To enhance model generalization, we show how our model and human confidence scores can be successfully applied to curriculum learning, a training strategy inspired by learning of `easier to harder' tasks. A higher model or human confidence score indicates a more recognisable and therefore easier sample, and can therefore be used as a scoring function to rank samples in curriculum learning. We evaluate our proposed methods with four state-of-the-art architectures for image and text classification task, using datasets with multi-rater label annotations by humans. We report that integrating model or human confidence information in label smoothing and curriculum learning improves both model performance and model calibration. The code are available at \url{https://github.com/AoShuang92/Confidence_Calibration_CL}.

Shuang Ao

Although AI systems have been applied in various fields and achieved impressive performance, their safety and reliability are still a big concern. This is especially important for safety-critical tasks. One shared characteristic of these critical tasks is their risk sensitivity, where small mistakes can cause big consequences and even endanger life. There are several factors that could be guidelines for the successful deployment of AI systems in sensitive tasks: (i) failure detection and out-of-distribution (OOD) detection; (ii) overfitting identification; (iii) uncertainty quantification for predictions; (iv) robustness to data perturbations. These factors are also challenges of current AI systems, which are major blocks for building safe and reliable AI. Specifically, the current AI algorithms are unable to identify common causes for failure detection. Furthermore, additional techniques are required to quantify the quality of predictions. All these contribute to inaccurate uncertainty quantification, which lowers trust in predictions. Hence obtaining accurate model uncertainty quantification and its further improvement are challenging. To address these issues, many techniques have been proposed, such as regularization methods and learning strategies. As vision and language are the most typical data type and have many open source benchmark datasets, this thesis will focus on vision-language data processing for tasks like classification, image captioning, and vision question answering. In this thesis, we aim to build a safeguard by further developing current techniques to ensure the accurate model uncertainty for safety-critical tasks.

Shuang Ao, Yi Dong, Jinwei Hu et al.

Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enhances adaptability while reducing computational costs. However, fine-tuning can compromise safety alignment, even with benign data, increasing susceptibility to harmful outputs. Existing safety alignment methods struggle to capture complex parameter shifts, leading to suboptimal safety-utility trade-offs. To address this issue, we propose Safe Pruning LoRA (SPLoRA), a novel pruning-based approach that selectively removes LoRA layers that weaken safety alignment, improving safety while preserving performance. At its core, we introduce Empirical-DIEM (E-DIEM), a dimension-insensitive similarity metric that effectively detects safety misalignment in LoRA-adapted models. We conduct extensive experiments on LLMs fine-tuned with mixed of benign and malicious data, and purely benign datasets, evaluating SPLoRA across utility, safety, and reliability metrics. Results demonstrate that SPLoRA outperforms state-of-the-art safety alignment techniques, significantly reducing safety risks while maintaining or improving model performance and reliability. Additionally, SPLoRA reduces inference overhead, making it a scalable and efficient solution for deploying safer and more reliable LLMs. The code is available at https://github.com/AoShuang92/SPLoRA.

Lihuan Li, Hao Xue, Shuang Ao et al.

The representation of urban trajectory data plays a critical role in effectively analyzing spatial movement patterns. Despite considerable progress, the challenge of designing trajectory representations that can capture diverse and complementary information remains an open research problem. Existing methods struggle in incorporating trajectory fine-grained details and high-level summary in a single model, limiting their ability to attend to both long-term dependencies while preserving local nuances. To address this, we propose HiT-JEPA (Hierarchical Interactions of Trajectory Semantics via a Joint Embedding Predictive Architecture), a unified framework for learning multi-scale urban trajectory representations across semantic abstraction levels. HiT-JEPA adopts a three-layer hierarchy that progressively captures point-level fine-grained details, intermediate patterns, and high-level trajectory abstractions, enabling the model to integrate both local dynamics and global semantics in one coherent structure. Extensive experiments on multiple real-world datasets for trajectory similarity computation show that HiT-JEPA's hierarchical design yields richer, multi-scale representations. Code is available at: https://anonymous.4open.science/r/HiT-JEPA.

Shuang Ao, Stefan Rueger, Advaith Siddharthan

Despite the impressive capability of large language models (LLMs), knowing when to trust their generations remains an open challenge. The recent literature on uncertainty quantification of natural language generation (NLG) utilises a conventional natural language inference (NLI) classifier to measure the semantic dispersion of LLMs responses. These studies employ logits of NLI classifier for semantic clustering to estimate uncertainty. However, logits represent the probability of the predicted class and barely contain feature information for potential clustering. Alternatively, CLIP (Contrastive Language-Image Pre-training) performs impressively in extracting image-text pair features and measuring their similarity. To extend its usability, we propose Contrastive Semantic Similarity, the CLIP-based feature extraction module to obtain similarity features for measuring uncertainty for text pairs. We apply this method to selective NLG, which detects and rejects unreliable generations for better trustworthiness of LLMs. We conduct extensive experiments with three LLMs on several benchmark question-answering datasets with comprehensive evaluation metrics. Results show that our proposed method performs better in estimating reliable responses of LLMs than comparable baselines. Results show that our proposed method performs better in estimating reliable responses of LLMs than comparable baselines. The code are available at \url{https://github.com/AoShuang92/css_uq_llms}.

Shuang Ao, Gopal Rumchurn

Adapting Large Language Models (LLMs) using parameter-efficient fine-tuning (PEFT) techniques such as LoRA has enabled powerful capabilities in LLM-based agents. However, these adaptations can unintentionally compromise safety alignment, leading to unsafe or unstable behaviors, particularly in agent planning tasks. Existing safety-aware adaptation methods often require access to both base and instruction-tuned model checkpoints, which are frequently unavailable in practice, limiting their applicability. We propose S3LoRA (Safe Spectral Sharpness-Guided Pruning LoRA), a lightweight, data-free, and model-independent framework that mitigates safety risks in LoRA-adapted models by inspecting only the fine-tuned weight updates. We first introduce Magnitude-Aware Spherically Normalized SVD (MAS-SVD), which robustly analyzes the structural properties of LoRA updates while preserving global magnitude information. We then design the Spectral Sharpness Index (SSI), a sharpness-aware metric to detect layers with highly concentrated and potentially unsafe updates. These layers are pruned post-hoc to reduce risk without sacrificing task performance. Extensive experiments and ablation studies across agent planning and language generation tasks show that S3LoRA consistently improves safety metrics while maintaining or improving utility metrics and significantly reducing inference cost. These results establish S3LoRA as a practical and scalable solution for safely deploying LLM-based agents in real-world, resource-constrained, and safety-critical environments.

Shuang Ao, Simon Khan, Haris Aziz et al.

Understanding the agent's learning process, particularly the factors that contribute to its success or failure post-training, is crucial for comprehending the rationale behind the agent's decision-making process. Prior methods clarify the learning process by creating a structural causal model (SCM) or visually representing the distribution of value functions. Nevertheless, these approaches have constraints as they exclusively function in 2D-environments or with uncomplicated transition dynamics. Understanding the agent's learning process in complicated environments or tasks is more challenging. In this paper, we propose REVEAL-IT, a novel framework for explaining the learning process of an agent in complex environments. Initially, we visualize the policy structure and the agent's learning process for various training tasks. By visualizing these findings, we can understand how much a particular training task or stage affects the agent's performance in test. Then, a GNN-based explainer learns to highlight the most important section of the policy, providing a more clear and robust explanation of the agent's learning process. The experiments demonstrate that explanations derived from this framework can effectively help in the optimization of the training tasks, resulting in improved learning efficiency and final performance.

Shuang Ao, Xeno Acharya

A medical dialogue system is essential for healthcare service as providing primary clinical advice and diagnoses. It has been gradually adopted and practiced in medical organizations in the form of a conversational bot, largely due to the advancement of NLP. In recent years, the introduction of state-of-the-art deep learning models and transfer learning techniques like Universal Language Model Fine Tuning (ULMFiT) and Knowledge Distillation (KD) largely contributes to the performance of NLP tasks. However, some deep neural networks are poorly calibrated and wrongly estimate the uncertainty. Hence the model is not trustworthy, especially in sensitive medical decision-making systems and safety tasks. In this paper, we investigate the well-calibrated model for ULMFiT and self-distillation (SD) in a medical dialogue system. The calibrated ULMFiT (CULMFiT) is obtained by incorporating label smoothing (LS), a commonly used regularization technique to achieve a well-calibrated model. Moreover, we apply the technique to recalibrate the confidence score called temperature scaling (TS) with KD to observe its correlation with network calibration. To further understand the relation between SD and calibration, we use both fixed and optimal temperatures to fine-tune the whole model. All experiments are conducted on the consultation backpain dataset collected by experts then further validated using a large publicly medial dialogue corpus. We empirically show that our proposed methodologies outperform conventional methods in terms of accuracy and robustness.