Context: Ensuring high levels of dependability in modern computer-based systems has become increasingly challenging due to their complexity. Although systems are validated at design time, their behavior can be different at run-time, possibly showing control-flow anomalies due to "unknown unknowns". Objective: We aim to detect control-flow anomalies through software monitoring, which verifies run-time behavior by logging software execution and detecting deviations from expected control flow. Methods: We propose a methodology to develop software monitors for control-flow anomaly detection through Large Language Models (LLMs) and conformance checking. The methodology builds on existing software development practices to maintain traditional V&V while providing an additional level of robustness and trustworthiness. It leverages LLMs to link design-time models and implementation code, automating source-code instrumentation. The resulting event logs are analyzed via conformance checking, an explainable and effective technique for control-flow anomaly detection. Results: We test the methodology on a case-study scenario from the European Railway Traffic Management System / European Train Control System (ERTMS/ETCS), which is a railway standard for modern interoperable railways. The results obtained from the ERTMS/ETCS case study demonstrate that LLM-based source-code instrumentation can achieve up to 84.775% control-flow coverage of the reference design-time process model, while the subsequent conformance checking-based anomaly detection reaches a peak performance of 96.610% F1-score and 93.515% AUC. Conclusion: Incorporating domain-specific knowledge to guide LLMs in source-code instrumentation significantly allowed obtaining reliable and quality software logs and enabled effective control-flow anomaly detection through conformance checking.

Automated emotion detection is widely used in applications ranging from well-being monitoring to high-stakes domains like mental health and hiring. However, models often rely on annotations that reflect dominant cultural norms, limiting model ability to recognize emotional expression in dialects often excluded from training data distributions, such as African American Vernacular English (AAVE). This study examines emotion recognition model performance on AAVE compared to General American English (GAE). We analyze 2.7 million tweets geo-tagged within Los Angeles. Texts are scored for strength of AAVE using computational approximations of dialect features. Annotations of emotion presence and intensity are collected on a dataset of 875 tweets with both high and low AAVE densities. To assess model accuracy on a task as subjective as emotion perception, we calculate community-informed "silver" labels where AAVE-dense tweets are labeled by African American, AAVE-fluent (ingroup) annotators. On our labeled sample, GPT and BERT-based models exhibit false positive prediction rates of anger on AAVE more than double than on GAE. SpanEmo, a popular text-based emotion model, increases false positive rates of anger from 25 percent on GAE to 60 percent on AAVE. Additionally, a series of linear regressions reveals that models and non-ingroup annotations are significantly more correlated with profanity-based AAVE features than ingroup annotations. Linking Census tract demographics, we observe that neighborhoods with higher proportions of African American residents are associated with higher predictions of anger (Pearson's correlation r = 0.27) and lower joy (r = -0.10). These results find an emergent safety issue of emotion AI reinforcing racial stereotypes through biased emotion classification. We emphasize the need for culturally and dialect-informed affective computing systems.

Large Language Models (LLMs) are increasingly deployed in safety-critical domains, yet remain susceptible to hallucinations. While prior works have proposed confidence representation methods for hallucination detection, most of these approaches rely on computationally expensive sampling strategies and often disregard the distinction between hallucination types. In this work, we introduce a principled evaluation framework that differentiates between extrinsic and intrinsic hallucination categories and evaluates detection performance across a suite of curated benchmarks. In addition, we leverage a recent attention-based uncertainty quantification algorithm and propose novel attention aggregation strategies that improve both interpretability and hallucination detection performance. Our experimental findings reveal that sampling-based methods like Semantic Entropy are effective for detecting extrinsic hallucinations but generally fail on intrinsic ones. In contrast, our method, which aggregates attention over input tokens, is better suited for intrinsic hallucinations. These insights provide new directions for aligning detection strategies with the nature of hallucination and highlight attention as a rich signal for quantifying model uncertainty.

Long context LLMs are vulnerable to prompt injection, where an attacker can inject an instruction in a long context to induce an LLM to generate an attacker-desired output. Existing prompt injection defenses are designed for short contexts. When extended to long-context scenarios, they have limited effectiveness. The reason is that an injected instruction constitutes only a very small portion of a long context, making the defense very challenging. In this work, we propose PISanitizer, which first pinpoints and sanitizes potential injected tokens (if any) in a context before letting a backend LLM generate a response, thereby eliminating the influence of the injected instruction. To sanitize injected tokens, PISanitizer builds on two observations: (1) prompt injection attacks essentially craft an instruction that compels an LLM to follow it, and (2) LLMs intrinsically leverage the attention mechanism to focus on crucial input tokens for output generation. Guided by these two observations, we first intentionally let an LLM follow arbitrary instructions in a context and then sanitize tokens receiving high attention that drive the instruction-following behavior of the LLM. By design, PISanitizer presents a dilemma for an attacker: the more effectively an injected instruction compels an LLM to follow it, the more likely it is to be sanitized by PISanitizer. Our extensive evaluation shows that PISanitizer can successfully prevent prompt injection, maintain utility, outperform existing defenses, is efficient, and is robust to optimization-based and strong adaptive attacks. The code is available at https://github.com/sleeepeer/PISanitizer.

While generative models have become increasingly prevalent across various domains, fundamental concerns regarding their reliability persist. A crucial yet understudied aspect of these models is the uncertainty quantification surrounding their distribution approximation capabilities. Current evaluation methodologies focus predominantly on measuring the closeness between the learned and the target distributions, neglecting the inherent uncertainty in these measurements. In this position paper, we formalize the problem of uncertainty quantification in generative model learning. We discuss potential research directions, including the use of ensemble-based precision-recall curves. Our preliminary experiments on synthetic datasets demonstrate the effectiveness of aggregated precision-recall curves in capturing model approximation uncertainty, enabling systematic comparison among different model architectures based on their uncertainty characteristics.

Cochlear implants (CI) significantly improve spoken language in children with severe-to-profound sensorineural hearing loss (SNHL), yet outcomes remain more variable than in children with normal hearing. This variability cannot be reliably predicted for individual children using age at implantation or residual hearing. This study aims to compare the accuracy of traditional machine learning (ML) to deep transfer learning (DTL) algorithms to predict post-CI spoken language development of children with bilateral SNHL using a binary classification model of high versus low language improvers. A total of 278 implanted children enrolled from three centers. The accuracy, sensitivity and specificity of prediction models based upon brain neuroanatomic features using traditional ML and DTL learning. DTL prediction models using bilinear attention-based fusion strategy achieved: accuracy of 92.39% (95% CI, 90.70%-94.07%), sensitivity of 91.22% (95% CI, 89.98%-92.47%), specificity of 93.56% (95% CI, 90.91%-96.21%), and area under the curve (AUC) of 0.977 (95% CI, 0.969-0.986). DTL outperformed traditional ML models in all outcome measures. DTL was significantly improved by direct capture of discriminative and task-specific information that are advantages of representation learning enabled by this approach over ML. The results support the feasibility of a single DTL prediction model for language prediction of children served by CI programs worldwide.

Guard models are a critical component of LLM safety, but their sensitivity to superficial linguistic variations remains a key vulnerability. We show that even meaning-preserving paraphrases can cause large fluctuations in safety scores, revealing a lack of semantic grounding. To address this, we introduce a practical, self-supervised framework for improving the semantic robustness of guard models. Our method leverages paraphrase sets to enforce prediction consistency using a novel, skew-aware aggregation strategy for robust target computation. Notably, we find that standard aggregation methods like mean and median can degrade safety, underscoring the need for skew-aware alternatives. We analyze six open-source guard models and show that our approach reduces semantic variability across paraphrases by ~58%, improves benchmark accuracy by ~2.5% on average, and generalizes to unseen stylistic variations. Intriguingly, we discover a bidirectional relationship between model calibration and consistency: our robustness training improves calibration by up to 40%, revealing a fundamental connection between these properties. These results highlight the value of treating semantic consistency as a first-class training objective and provide a scalable recipe for building more reliable guard models.

Large language models (LLMs) are increasingly used to extract structured information from free-text clinical records, but prior work often focuses on single tasks, limited models, and English-language reports. We evaluated 15 open-weight LLMs on pathology and radiology reports across six use cases, colorectal liver metastases, liver tumours, neurodegenerative diseases, soft-tissue tumours, melanomas, and sarcomas, at three institutes in the Netherlands, UK, and Czech Republic. Models included general-purpose and medical-specialised LLMs of various sizes, and six prompting strategies were compared: zero-shot, one-shot, few-shot, chain-of-thought, self-consistency, and prompt graph. Performance was assessed using task-appropriate metrics, with consensus rank aggregation and linear mixed-effects models quantifying variance. Top-ranked models achieved macro-average scores close to inter-rater agreement across tasks. Small-to-medium general-purpose models performed comparably to large models, while tiny and specialised models performed worse. Prompt graph and few-shot prompting improved performance by ~13%. Task-specific factors, including variable complexity and annotation variability, influenced results more than model size or prompting strategy. These findings show that open-weight LLMs can extract structured data from clinical reports across diseases, languages, and institutions, offering a scalable approach for clinical data curation.

When deploying LLMs in agentic architectures requiring real-time decisions under temporal constraints, we assume they reliably determine whether action windows remain open or have closed. This assumption is untested. We characterize temporal constraint processing across eight production-scale models (2.8-8B parameters) using deadline detection tasks, revealing systematic deployment risks: bimodal performance distribution (models achieve either 95% or 50% accuracy), extreme prompt brittleness (30-60 percentage point swings from formatting changes alone), and systematic action bias (100% false positive rates in failing models). Parameter count shows no correlation with capability in this range-a 3.8B model matches 7B models while other 7B models fail completely. Fine-tuning on 200 synthetic examples improves models with partial capability by 12-37 percentage points. We demonstrate that temporal constraint satisfaction cannot be reliably learned through next-token prediction on natural language, even with targeted fine-tuning. This capability requires architectural mechanisms for: (1) continuous temporal state representation, (2) explicit constraint checking separate from linguistic pattern matching, (3) systematic compositional reasoning over temporal relations. Current autoregressive architectures lack these mechanisms. Deploying such systems in time-critical applications without hybrid architectures incorporating symbolic reasoning modules represents unacceptable risk.

Efficient retrieval of external knowledge bases and web pages is crucial for enhancing the reasoning abilities of LLMs. Previous works on training LLMs to leverage external retrievers for solving complex problems have predominantly employed end-to-end reinforcement learning. However, these approaches neglect supervision over the reasoning process, making it difficult to guarantee logical coherence and rigor. To address these limitations, we propose Thinker, a hierarchical thinking model for deep search through multi-turn interaction, making the reasoning process supervisable and verifiable. It decomposes complex problems into independently solvable sub-problems, each dually represented in both natural language and an equivalent logical function to support knowledge base and web searches. Concurrently, dependencies between sub-problems are passed as parameters via these logical functions, enhancing the logical coherence of the problem-solving process. To avoid unnecessary external searches, we perform knowledge boundary determination to check if a sub-problem is within the LLM's intrinsic knowledge, allowing it to answer directly. Experimental results indicate that with as few as several hundred training samples, the performance of Thinker is competitive with established baselines. Furthermore, when scaled to the full training set, Thinker significantly outperforms these methods across various datasets and model sizes. The source code is available at https://github.com/OpenSPG/KAG-Thinker.