Activation steering has emerged as a powerful tool to shape LLM behaviour without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors obtained with Contrastive Activation Addition (CAA), a widely used steering approach, under a unified evaluation protocol. We show that steering vectors consistently influence the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. Across LLM families and sizes, steering the model in specific directions can drastically increase (by up to 57%) or decrease (by up to 50%) its attack success rate (ASR), depending on the targeted behaviour. We attribute this phenomenon to the overlap between the steering vectors and the latent subspace of refusal behaviour. Thus, we offer a mechanistic explanation for this discovery. Together, our findings reveal the previously unobserved origin of this safety gap in LLMs, highlighting a trade-off between controllability and safety. We release our code at https://github.com/yetiiil/analyse-sv-safety.

Generating high-fidelity synthetic tabular data remains a critical challenge for enhancing data availability in privacy-sensitive and low-resource domains. Recent approaches leverage LLMs by representing table rows as sequences, yet suffer from two fundamental limitations: (1) they model feature dependencies densely, introducing spurious correlations; and (2) they assume static relationships between features, ignoring how these dependencies vary with feature values. To overcome these limitations, we introduce SAGE (Sparse Adaptive Guidance), a novel LLM-based generation framework that enforces sparse and dynamic dependency guidance. SAGE discretizes features into value-aware pseudo-features and constructs a mutual information-based sparse dependency graph. This graph adaptively guides generation through explicit context selection or implicit logit correction, enabling LLMs to focus on truly relevant information during synthesis. Our extensive experiments across six datasets and multiple tasks reveal that SAGE not only improves data fidelity and downstream utility, boosting F1 scores by 10% compared to previous LLM-based methods, but also reduces policy violations by one point. These results highlight the importance of adaptive structure in tabular data generation and provide new insights into context-sensitive control of LLMs.

Ensuring the robustness of factual knowledge in LLMs is critical for reliable applications in tasks such as question answering and reasoning. However, existing evaluation methods predominantly focus on performance-based metrics, often investigating from the perspective of prompt perturbations, which captures only the externally triggered side of knowledge robustness. To bridge this gap, we introduce a principled approach to measure factual robustness from the perspective of the generation process by analyzing token distribution entropy in combination with temperature scaling sensitivity. These two factors build the Factual Robustness Score (FRS), a novel metric which quantifies the stability of a fact against perturbations in decoding conditions, given its initial uncertainty. To validate our approach, we conduct extensive experiments on 5 LLMs across 3 closed-book QA datasets (SQuAD, TriviaQA, and HotpotQA). We show that factual robustness varies significantly – smaller models report an FRS of 0.76, larger ones 0.93 – with accuracy degrading by ~60% under increased uncertainty. These insights demonstrate how entropy and temperature scaling impact factual accuracy, and lay a foundation for developing more robust knowledge retention and retrieval in future models. We release our code at https://github.com/afastowski/frs.

Large Language Models (LLMs) have shown strong potential for tabular data generation by modeling textualized feature-value pairs. However, tabular data inherently exhibits sparse feature-level dependencies, where many feature interactions are structurally insignificant. This creates a fundamental mismatch as LLMs’ self-attention mechanism inevitably distributes focus across all pairs, diluting attention on critical relationships, particularly in datasets with complex dependencies or semantically ambiguous features. To address this limitation, we propose GraDe (Graph-Guided Dependency Learning), a novel method that explicitly integrates sparse dependency graphs into LLMs’ attention mechanism. GraDe employs a lightweight dynamic graph learning module guided by externally extracted functional dependencies, prioritizing key feature interactions while suppressing irrelevant ones. Our experiments across diverse real-world datasets demonstrate that GraDe outperforms existing LLM-based approaches by up to 12% on complex datasets while achieving competitive results with state-of-the-art approaches in synthetic data quality. Our method is minimally intrusive yet effective, offering a practical solution for structure-aware tabular data modeling with LLMs.

Pre-trained language models have achieved remarkable success across diverse applications but remain susceptible to spurious, concept-driven correlations that impair robustness and fairness. In this work, we introduce CURE, a novel and lightweight framework that systematically disentangles and suppresses conceptual shortcuts while preserving essential content information. Our method first extracts concept-irrelevant representations via a dedicated content extractor reinforced by a reversal network, ensuring minimal loss of task-relevant information. A subsequent controllable debiasing module employs contrastive learning to finely adjust the influence of residual conceptual cues, enabling the model to either diminish harmful biases or harness beneficial correlations as appropriate for the target task. Evaluated on the IMDB and Yelp datasets using three pre-trained architectures, CURE achieves an absolute improvement of +10 points in F1 score on IMDB and +2 points on Yelp, while introducing minimal computational overhead. Our approach establishes a flexible, unsupervised blueprint for combating conceptual biases, paving the way for more reliable and fair language understanding systems.

Tabular data is critical across diverse domains, yet high-quality datasets remain scarce due to privacy concerns and the cost of collection. Contemporary approaches adopt large language models (LLMs) for tabular augmentation, but exhibit two major limitations: (1) dense dependency modeling among tabular features that can introduce bias, and (2) high computational overhead in sampling. To address these issues, we propose SPADA for SPArse Dependency-driven Augmentation, a lightweight generative framework that explicitly captures sparse dependencies via an LLM-induced graph. We treat each feature as a node and synthesize values by traversing the graph, conditioning each feature solely on its parent nodes. We explore two synthesis strategies: a non-parametric method using Gaussian kernel density estimation, and a conditional normalizing flow model that learns invertible mappings for conditional density estimation. Experiments on four datasets show that SPADA reduces constraint violations by 4% compared to diffusion-based methods and accelerates generation by nearly 9,500× over LLM-based baselines.

Large Language Models (LLMs) have shown impressive moral reasoning abilities. Yet they often diverge when confronted with complex, multi-factor moral dilemmas. To address these discrepancies, we propose a framework that synthesizes multiple LLMs’ moral judgments into a collectively formulated moral judgment, realigning models that deviate significantly from this consensus. Our aggregation mechanism fuses continuous moral acceptability scores (beyond binary labels) into a collective probability, weighting contributions by model reliability. For misaligned models, a targeted embedding-optimization procedure fine-tunes token embeddings for moral philosophical theories, minimizing JS divergence to the consensus while preserving semantic integrity. Experiments on a large-scale social moral dilemma dataset show our approach builds robust consensus and improves individual model fidelity. These findings highlight the value of data-driven moral alignment across multiple models and its potential for safer, more consistent AI systems.