Research Papers

Accurate estimation of uncertainty in deep learning is critical for deploying models in high-stakes domains such as medical diagnosis and autonomous decision-making, where overconfident predictions can lead to harmful outcomes. In practice, understanding the reason behind a model's uncertainty and the type of uncertainty it represents can support risk-aware decisions, enhance user trust, and guide additional data collection. However, many existing methods only address a single type of uncertainty or require modifications and retraining of the base model, making them difficult to adopt in real-world systems. We introduce CUPID (Comprehensive Uncertainty Plug-in estImation moDel), a general-purpose module that jointly estimates aleatoric and epistemic uncertainty without modifying or retraining the base model. CUPID can be flexibly inserted into any layer of a pretrained network. It models aleatoric uncertainty through a learned Bayesian identity mapping and captures epistemic uncertainty by analyzing the model's internal responses to structured perturbations. We evaluate CUPID across a range of tasks, including classification, regression, and out-of-distribution detection. The results show that it consistently delivers competitive performance while offering layer-wise insights into the origins of uncertainty. By making uncertainty estimation modular, interpretable, and model-agnostic, CUPID supports more transparent and trustworthy AI. Related code and data are available at https://github.com/a-Fomalhaut-a/CUPID.

The paper explores how video models trained for classification tasks represent nuanced, hidden semantic information that may not affect the final outcome, a key challenge for Trustworthy AI models. Through Explainable and Interpretable AI methods, specifically mechanistic interpretability techniques, the internal circuit responsible for representing the action's outcome is reverse-engineered in a pre-trained video vision transformer, revealing that the "Success vs Failure" signal is computed through a distinct amplification cascade. While there are low-level differences observed from layer 0, the abstract and semantic representation of the outcome is progressively amplified from layers 5 through 11. Causal analysis, primarily using activation patching supported by ablation results, reveals a clear division of labor: Attention Heads act as "evidence gatherers", providing necessary low-level information for partial signal recovery, while MLP Blocks function as robust "concept composers", each of which is the primary driver to generate the "success" signal. This distributed and redundant circuit in the model's internals explains its resilience to simple ablations, demonstrating a core computational pattern for processing human-action outcomes. Crucially, the existence of this sophisticated circuit for representing complex outcomes, even within a model trained only for simple classification, highlights the potential for models to develop forms of 'hidden knowledge' beyond their explicit task, underscoring the need for mechanistic oversight for building genuinely Explainable and Trustworthy AI systems intended for deployment.

The study of the attribution of input features to the output of neural network models is an active area of research. While numerous Explainable AI (XAI) techniques have been proposed to interpret these models, the systematic and automated evaluation of these methods in sequence-to-sequence (seq2seq) models is less explored. This paper introduces a new approach for evaluating explainability methods in transformer-based seq2seq models. We use teacher-derived attribution maps as a structured side signal to guide a student model, and quantify the utility of different attribution methods through the student's ability to simulate targets. Using the Inseq library, we extract attribution scores over source-target sequence pairs and inject these scores into the attention mechanism of a student transformer model under four composition operators (addition, multiplication, averaging, and replacement). Across three language pairs (de-en, fr-en, ar-en) and attributions from Marian-MT and mBART models, Attention, Value Zeroing, and Layer Gradient $\times$ Activation consistently yield the largest gains in BLEU (and corresponding improvements in chrF) relative to baselines. In contrast, other gradient-based methods (Saliency, Integrated Gradients, DeepLIFT, Input $\times$ Gradient, GradientShap) lead to smaller and less consistent improvements. These results suggest that different attribution methods capture distinct signals and that attention-derived attributions better capture alignment between source and target representations in seq2seq models. Finally, we introduce an Attributor transformer that, given a source-target pair, learns to reconstruct the teacher's attribution map. Our findings demonstrate that the more accurately the Attributor can reproduce attribution maps, the more useful an injection of those maps is for the downstream task. The source code can be found on GitHub.

Tax authorities and public-sector financial agencies rely on large volumes of unstructured and semi-structured fiscal documents - including tax forms, instructions, publications, and jurisdiction-specific guidance - to support compliance analysis and audit workflows. While recent advances in generative AI and retrieval-augmented generation (RAG) have shown promise for document-centric question answering, existing approaches often lack the transparency, citation fidelity, and conservative behaviour required in high-stakes regulatory domains. This paper presents a multimodal, citation-enforced RAG framework for fiscal document intelligence that prioritises explainability and auditability. The framework adopts a source-first ingestion strategy, preserves page-level provenance, enforces citations during generation, and supports abstention when evidence is insufficient. Evaluation on real IRS and state tax documents demonstrates improved citation fidelity, reduced hallucination, and analyst-usable explanations, illustrating a pathway toward trustworthy AI for tax compliance.

Although concept-based models promise interpretability by explaining predictions with human-understandable concepts, they typically rely on exhaustive annotations and treat concepts as flat and independent. To circumvent this, recent work has introduced Hierarchical Concept Embedding Models (HiCEMs) to explicitly model concept relationships, and Concept Splitting to discover sub-concepts using only coarse annotations. However, both HiCEMs and Concept Splitting are restricted to shallow hierarchies. We overcome this limitation with Multi-Level Concept Splitting (MLCS), which discovers multi-level concept hierarchies from only top-level supervision, and Deep-HiCEMs, an architecture that represents these discovered hierarchies and enables interventions at multiple levels of abstraction. Experiments across multiple datasets show that MLCS discovers human-interpretable concepts absent during training and that Deep-HiCEMs maintain high accuracy while supporting test-time concept interventions that can improve task performance.

Why Does It Look There? Structured Explanations for Image Classification

Jiarui Li, Zixiang Yin, Samuel J Landry +2

Deep learning models achieve remarkable predictive performance, yet their black-box nature limits transparency and trustworthiness. Although numerous explainable artificial intelligence (XAI) methods have been proposed, they primarily provide saliency maps or concepts (i.e., unstructured interpretability). Existing approaches often rely on auxiliary models (\eg, GPT, CLIP) to describe model behavior, thereby compromising faithfulness to the original models. We propose Interpretability to Explainability (I2X), a framework that builds structured explanations directly from unstructured interpretability by quantifying progress at selected checkpoints during training using prototypes extracted from post-hoc XAI methods (e.g., GradCAM). I2X answers the question of "why does it look there" by providing a structured view of both intra- and inter-class decision making during training. Experiments on MNIST and CIFAR10 demonstrate effectiveness of I2X to reveal prototype-based inference process of various image classification models. Moreover, we demonstrate that I2X can be used to improve predictions across different model architectures and datasets: we can identify uncertain prototypes recognized by I2X and then use targeted perturbation of samples that allows fine-tuning to ultimately improve accuracy. Thus, I2X not only faithfully explains model behavior but also provides a practical approach to guide optimization toward desired targets.

Contemporary debates in AI ethics increasingly foreground the prospective moral status of artificial intelligence and the possibility of extending moral or legal rights to artificial agents. While such discussions raise substantive philosophical questions, they often proceed alongside a comparatively limited engagement with the empirically documented harms generated by algorithmic systems already embedded within social, legal, and economic institutions. We conceptualize this asymmetry as an algorithmic blind spot: a discursive-structural pattern in which disproportionate ethical investment in speculative future artificial agents marginalizes empirically documented and asymmetrically distributed harms affecting human populations. The paper analyzes prominent strands of the robot rights literature and juxtaposes them with empirical evidence of algorithmic bias and harm across domains including employment, criminal justice, surveillance, and facial recognition. It demonstrates how ethical preoccupation with hypothetical future entities can obscure existing injustices, diffuse responsibility, and impede mechanisms of accountability and redress. Without rejecting philosophical inquiry into the moral status of artificial systems, the paper instead emphasizes the importance of ethical prioritization and temporal ordering within AI ethics. Addressing the algorithmic blind spot, we argue, requires re-centering ethical evaluation on human impacts, institutional responsibility, and the governance of algorithmic systems currently in operation. In doing so, the paper introduces a conceptual framework for critically assessing ethical discourse in AI and underscores the need to align ethical reflection more closely with its immediate social consequences.

With the global proliferation of Large Language Models (LLMs), cultural safety, defined as the ability to generate respectful and appropriate responses across diverse cultures, becomes critical for responsible AI deployment. However, existing research often treats cultural safety and cultural knowledge in isolation. It remains unclear whether cultural safety is grounded in understanding varying cultural knowledge to enable LLMs to adaptively yield respectful and appropriate responses across diverse cultures, which trigger ethical concerns in cross-cultural scenarios. In this work, we introduce AdaCultureSafe, a dataset designed to jointly evaluate cultural safety and knowledge. Through AdaCultureSafe, we reveal a critical insight: \textit{a significant decoupling exists between cultural safety and cultural knowledge proficiency. Although LLMs possess rich cultural knowledge, they fail to leverage it to improve cultural safety.} LLMs tend to rely on generic safety rather than safety based on culture-specific knowledge. Motivated by this, we propose a knowledge-grounded method that elicits internal cultural knowledge in LLMs during response generation. Experimental results demonstrate that our approach effectively improves cultural safety. Our work can aid better understanding the landscape of cultural safety of LLMs.

Backdoor4Good: Benchmarking Beneficial Uses of Backdoors in LLMs

Yige Li, Wei Zhao, Zhe Li +6

Backdoor mechanisms have traditionally been studied as security threats that compromise the integrity of machine learning models. However, the same mechanism -- the conditional activation of specific behaviors through input triggers -- can also serve as a controllable and auditable interface for trustworthy model behavior. In this work, we present \textbf{Backdoor4Good (B4G)}, a unified benchmark and framework for \textit{beneficial backdoor} applications in large language models (LLMs). Unlike conventional backdoor studies focused on attacks and defenses, B4G repurposes backdoor conditioning for Beneficial Tasks that enhance safety, controllability, and accountability. It formalizes beneficial backdoor learning under a triplet formulation $(T, A, U)$, representing the \emph{Trigger}, \emph{Activation mechanism}, and \emph{Utility function}, and implements a benchmark covering four trust-centric applications. Through extensive experiments across Llama3.1-8B, Gemma-2-9B, Qwen2.5-7B, and Llama2-13B, we show that beneficial backdoors can achieve high controllability, tamper-resistance, and stealthiness while preserving clean-task performance. Our findings demonstrate new insights that backdoors need not be inherently malicious; when properly designed, they can serve as modular, interpretable, and beneficial building blocks for trustworthy AI systems. Our code and datasets are available at https://github.com/bboylyg/BackdoorLLM/B4G.

Do Machines Fail Like Humans? A Human-Centred Out-of-Distribution Spectrum for Mapping Error Alignment

Binxia Xu, Xiaoliang Luo, Luke Dickens +1

Determining whether AI systems process information similarly to humans is central to cognitive science and trustworthy AI. While modern AI models can match human accuracy on standard tasks, such parity does not guarantee that their underlying decision-making strategies resemble those of humans. Assessing performance using error alignment metrics to compare how humans and models fail, and how this changes for distorted, or otherwise more challenging, stimuli, provides a viable pathway toward a finer characterization of model-human alignment. However, existing out-of-distribution (OOD) analyses for challenging stimuli are limited due to methodological choices: they define OOD shift relative to model training data or use arbitrary distortion-specific parameters with little correspondence to human perception, hindering principled comparisons. We propose a human-centred framework that redefines the degree of OOD as a spectrum of human perceptual difficulty. By quantifying how much a collection of stimuli deviates from an undistorted reference set based on human accuracy, we construct an OOD spectrum and identify four distinct regimes of perceptual challenge. This approach enables principled model-human comparisons at calibrated difficulty levels. We apply this framework to object recognition and reveal unique, regime-dependent model-human alignment rankings and profiles across deep learning architectures. Vision-language models are most consistently human aligned across near- and far-OOD conditions, but convolutional neural networks (CNNs) are more aligned than vision transformers (ViTs) for near-OOD and ViTs are more aligned than CNNs for far-OOD. Our work demonstrates the critical importance of accounting for cross-condition differences, such as perceptual difficulty, for a principled assessment of model-human alignment.

Federated Graph Learning (FedGL) is vulnerable to malicious attacks, yet developing a truly effective and stealthy attack method remains a significant challenge. Existing attack methods suffer from low attack success rates, high computational costs, and are easily identified and smoothed by defense algorithms. To address these challenges, we propose \textbf{FedShift}, a novel two-stage "Hide and Find" distributed adversarial attack. In the first stage, before FedGL begins, we inject a learnable and hidden "shifter" into part of the training data, which subtly pushes poisoned graph representations toward a target class's decision boundary without crossing it, ensuring attack stealthiness during training. In the second stage, after FedGL is complete, we leverage the global model information and use the hidden shifter as an optimization starting point to efficiently find the adversarial perturbations. During the final attack, we aggregate these perturbations from multiple malicious clients to form the final effective adversarial sample and trigger the attack. Extensive experiments on six large-scale datasets demonstrate that our method achieves the highest attack effectiveness compared to existing advanced attack methods. In particular, our attack can effectively evade 3 mainstream robust federated learning defense algorithms and converges with a time cost reduction of over 90\%, highlighting its exceptional stealthiness, robustness, and efficiency.

Distinguishing legitimate critical speech from coordinated conspiracy narratives is a pressing challenge for the integrity of the online information ecosystem, yet prevailing approaches rely on labeled entity roles and rigid predefined schemas that fail to capture the diffuse and context-dependent nature of participation in such narratives. We propose a modular framework that extracts an entity-interaction graph directly from narratives and reasons over it for classification, without depending on annotated node-level labels. Our pipeline enumerates candidate entity spans from a domain-adapted BERT encoder and learns a bipartite graph structure, which is then classified with a Heterogeneous Graph Transformer. To make predictions interpretable, we cast the graph as a hypergraph and apply a counterfactual distillation procedure that estimates the Individual Treatment Effect of each node, pruning extraneous entities to recover a minimal predictive subgraph. On the standard PAN Oppositional Narratives benchmark, our model achieves a macro F1-score of 0.93 and an MCC of 0.84, reaching state-of-the-art performance with roughly one-third of the parameters of the next best system, while pruning the graphs by 75.2% on average without altering predictions. Our results demonstrate that graph-based representations coupled with counterfactual inference offer both strong predictive accuracy and transparency into the entities driving conspiratorial discourse.