Sparse Autoencoders (SAEs) decompose language model activations into sparse, interpretable features, but these features often struggle to predict task-relevant concepts that remain clearly encoded in the original dense activations. We call this mismatch a concept faithfulness gap: a faithful SAE should preserve concept-level representations, such as toxicity or sentiment, rather than producing features that look interpretable but lose predictive relevance. We introduce ProbeSAE, a probing-guided fine-tuning framework that addresses this gap by preserving task-relevant concept structure in a more monosemantic sparse decomposition, enabling more effective causal control over model behaviour without sacrificing interpretability. We evaluate ProbeSAE on two SAE architectures (TopK, JumpReLU), two domains (toxicity, sentiment), and multilingual out-of-distribution settings spanning six languages. Trained exclusively on English data, ProbeSAE: (i) produces more predictive, monosemantic, and interpretable sparse features than pre-trained or naïvely fine-tuned SAEs across all six evaluation languages; (ii) learns language-independent feature directions, as confirmed by cross-lingual causal ablation; (iii) passes SAEBench evaluation, confirming these gains do not arise from task/domain overfitting. Together, these results show that the concept faithfulness gap in SAEs can be addressed through probe-guided training, improving monosemanticity and causal effectiveness without sacrificing interpretability, making probe-guided SAEs a more reliable tool for domain-specific interpretability analysis.

Chain-of-Thought prompting is increasingly relied upon as a safety monitoring surface: if a model reasons before acting, its trace can be inspected before harm occurs. Yet reasoning is a stochastic branching process rather than a deterministic, auditable path. We introduce Bifurcation Entropy (H_BE), a step-level diagnostic metric that extends Kernel Language Entropy to intermediate reasoning steps, measuring the semantic breadth of a model’s next-step distribution at each point in the Chain-of-Thought. We deploy H_BE via In-Distribution Attacks, a controlled probing instrument that injects the model’s own high-entropy continuations to test whether H_BE predicts exploitability. Across five reasoning models in high-stakes blackmail scenarios, phrases in the top H_BE quartile are exploited 6.2× more often than the lowest, with H_BE-targeted probing inducing significantly more alignment failures than random selection. These results establish H_BE as a principled probing signal that concentrates safety monitoring effort on the minority of reasoning steps that carry the majority of alignment risk.

Ensuring the safety of Large Language Models (LLMs) is a critical alignment challenge. Existing approaches often rely on invasive fine- tuning or external generation-based checks, which can be opaque and resource-inefficient. In this work, we investigate the geometry of safety concepts within pretrained representations, proposing a mechanistic methodology that identifies the layer where safe and unsafe concepts are maximally separable within a pretrained model’s representation space. By leveraging the intrinsic activation space of the optimal layer, we show that safety enforcement can be achieved via a simple linear classifier, avoiding the need for weight modification. We validate our framework across multiple domains (regulation, law, finance, cybersecurity, education, code, human resources, and social media), diverse tasks (safety classification, prompt injection, and toxicity detection), and 16 non-English languages on both encoder and decoder architectures. Our results show that: (i) the separation between safe and unsafe concepts emerges from a single layer direction in the activation space, (ii) monitoring internal representations provides a significantly more robust safeguarding mechanism compared to traditional evaluative or generative guardrail paradigms.

In a rapidly evolving labor market, detecting and addressing emerging skill needs is essential for shaping responsive education and workforce policies. Online job advertisements (OJAs) provide a real-time view of changing demands, but require first retrieving skill mentions from unstructured text and then solving the entity linking problem of connecting them to standardized skill taxonomies. To harness this potential, we present a multilingual human-in-the-loop (HITL) pipeline that operates in two steps: candidate skills are extracted from national OJA corpora using country-specific word embeddings, capturing terms that reflect each country’s labor market. These candidates are linked to ESCO using an encoder-based system and refined through a decoder large language models (LLMs) for accurate contextual alignment. Our approach is validated through both quantitative and qualitative evaluations, demonstrating that our method enables timely, multilingual monitoring of emerging skills, supporting agile policy-making and targeted training initiatives.

We present ITALIC, a large-scale benchmark dataset of 10,000 multiple-choice questions designed to evaluate the natural language understanding of the Italian language and culture. ITALIC spans 12 domains, exploiting public tests to score domain experts in real-world scenarios. We detail our data collection process, stratification techniques, and selection strategies. ITALIC provides a comprehensive assessment suite that captures commonsense reasoning and linguistic proficiency in a morphologically rich language. We establish baseline performances using 17 state-of-the-art LLMs, revealing current limitations in Italian language understanding and highlighting significant linguistic complexity and cultural specificity challenges. ITALIC serves as a benchmark for evaluating existing models and as a roadmap for future research, encouraging the development of more sophisticated and culturally aware natural language systems.

The influence of personas on Large Language Models (LLMs) has been widely studied, yet their direct impact on performance remains uncertain. This work explores a novel approach to guiding LLM behaviour through role vectors, an alternative to persona-based prompting. We construct 29 role vectors derived from model activations and evaluate their impact on benchmark performance across multiple domains. Our analysis investigates whether these vectors can effectively steer models toward domain-specific expertise. We measure two key interventions: (i) activation addition, which reinforces role-specific directions, and (ii) directional ablation, which removes them. Results on well-established benchmarks indicate that role vectors do, in fact, influence model behaviour, improving in-domain task performance while also yielding unexpected cross-domain gains.This, in turn, suggests that manipulating internal model representations has a greater impact on outcomes than persona-based prompting.

Despite the state-of-the-art performance of Large Language Models (LLMs), these models often suffer from hallucinations, which can undermine their performance in critical applications. In this work, we propose SAFE, a novel framework for detecting and mitigating hallucinations by leveraging Sparse Autoencoders (SAEs). While hallucination detection techniques and SAEs have been explored independently, their synergistic application in a comprehensive system, particularly for hallucination-aware query enrichment, has not been fully investigated. To validate the effectiveness of SAFE, we evaluate it on two models with available SAEs across four diverse cross-domain datasets designed to assess hallucination problems. Empirical results demonstrate that SAFE consistently improves query generation accuracy and mitigates hallucinations across all datasets, achieving accuracy improvements of up to 29.45%.

Interpreting the internal representations of large language models (LLMs) is crucial for their deployment in real-world applications, impacting areas such as AI safety, debugging, and compliance. Sparse Autoencoders facilitate interpretability by decomposing polysemantic activation into a latent space of monosemantic features. However, evaluating the auto-interpretability of these features is difficult and computationally expensive, which limits scalability in practical settings. In this work, we propose SFAL, an alternative evaluation strategy that reduces reliance on LLM-based scoring by assessing the alignment between the semantic neighbourhoods of features (derived from auto-interpretation embeddings) and their functional neighbourhoods (derived from co-occurrence statistics).Our method enhances efficiency, enabling fast and cost-effective assessments. We validate our approach on large-scale models, demonstrating its potential to provide interpretability while reducing computational overhead, making it suitable for real-world deployment.

Enriching sentences with knowledge from qualitative sources benefits various NLP tasks and enhances the use of labeled data in model training. This is crucial for Financial Sentiment Analysis (FSA), where texts are often brief and contain implied information. We introduce RE-FIN (Retrieval-based Enrichment for FINancial data), an automated system designed to retrieve information from a knowledge base to enrich financial sentences, making them more knowledge-dense and explicit. RE-FIN generates propositions from the knowledge base and employs Retrieval-Augmented Generation (RAG) to augment the original text with relevant information. A large language model (LLM) rewrites the original sentence, incorporating this data. Since the LLM does not create new content, the risk of hallucinations is significantly reduced. The LLM generates multiple new sentences using different relevant information from the knowledge base; we developed an algorithm to select one that best preserves the meaning of the original sentence while avoiding excessive syntactic similarity. Results show that enhanced sentences present lower perplexity than the original ones and improve performances on FSA.

We present BEEP (BEst DrivEr’s License Performer), a benchmark challenge to evaluate large language models in the context of a simulated Italian driver’s license exam. This challenge tests the models’ ability to understand and apply traffic laws, road safety regulations, and vehicle-related knowledge through a series of true/false questions. The dataset is derived from official ministerial materials used in the Italian licensing process, specifically targeting Category B licenses.We evaluate models such as LLaMA and Mixtral across multiple categories. In addition, we simulate a driving license test to assess the models’ real-world applicability, where the pass rate is determined based on the number of errors allowed. While scaling up model size improved performance, even larger models struggled to pass the exam consistently. The challenge demonstrates the capabilities and limitations of LLMs in handling real-world, high-stakes scenarios, providing insights into their practical use and areas for further improvement.

The recent growth of black-box machine-learning methods in data analysis has increased the demand for explanation methods and tools to understand their behaviour and assist human-ML model cooperation. In this paper, we demonstrate ContrXT, a novel approach that uses natural language explanations to help users to comprehend how a back-box model works. ContrXT provides time contrastive (t-contrast) explanations by computing the differences in the classification logic of two different trained models and then reasoning on their symbolic representations through Binary Decision Diagrams. ContrXT is publicly available at ContrXT.ai as a python pip package.