George Turkiyyah

Faris Hijazi, Somayah AlHarbi, Abdulaziz AlHussein et al.

The rapid advancements in Large Language Models (LLMs) have led to significant improvements in various natural language processing tasks. However, the evaluation of LLMs' legal knowledge, particularly in non-English languages such as Arabic, remains under-explored. To address this gap, we introduce ArabLegalEval, a multitask benchmark dataset for assessing the Arabic legal knowledge of LLMs. Inspired by the MMLU and LegalBench datasets, ArabLegalEval consists of multiple tasks sourced from Saudi legal documents and synthesized questions. In this work, we aim to analyze the capabilities required to solve legal problems in Arabic and benchmark the performance of state-of-the-art LLMs. We explore the impact of in-context learning and investigate various evaluation methods. Additionally, we explore workflows for generating questions with automatic validation to enhance the dataset's quality. We benchmark multilingual and Arabic-centric LLMs, such as GPT-4 and Jais, respectively. We also share our methodology for creating the dataset and validation, which can be generalized to other domains. We hope to accelerate AI research in the Arabic Legal domain by releasing the ArabLegalEval dataset and code: https://github.com/Thiqah/ArabLegalEval

Stefano Zampini, Umberto Zerbinati, George Turkiyyah et al.

In recent years, we have witnessed the emergence of scientific machine learning as a data-driven tool for the analysis, by means of deep-learning techniques, of data produced by computational science and engineering applications. At the core of these methods is the supervised training algorithm to learn the neural network realization, a highly non-convex optimization problem that is usually solved using stochastic gradient methods. However, distinct from deep-learning practice, scientific machine-learning training problems feature a much larger volume of smooth data and better characterizations of the empirical risk functions, which make them suited for conventional solvers for unconstrained optimization. We introduce a lightweight software framework built on top of the Portable and Extensible Toolkit for Scientific computation to bridge the gap between deep-learning software and conventional solvers for unconstrained minimization. We empirically demonstrate the superior efficacy of a trust region method based on the Gauss-Newton approximation of the Hessian in improving the generalization errors arising from regression tasks when learning surrogate models for a wide range of scientific machine-learning techniques and test cases. All the conventional second-order solvers tested, including L-BFGS and inexact Newton with line-search, compare favorably, either in terms of cost or accuracy, with the adaptive first-order methods used to validate the surrogate models.

Harethah Abu Shairah, Hasan Abed Al Kader Hammoud, Bernard Ghanem et al.

Large language models (LLMs) are typically aligned to refuse harmful instructions through safety fine-tuning. A recent attack, termed abliteration, identifies and suppresses the single latent direction most responsible for refusal behavior, thereby enabling models to generate harmful content. We propose a defense that fundamentally alters how models express refusal. We construct an extended-refusal dataset in which responses to harmful prompts provide detailed justifications before refusing, distributing the refusal signal across multiple token positions. Fine-tuning Llama-2-7B-Chat and Qwen2.5-Instruct (1.5B and 3B parameters) on this dataset yields models that maintain high refusal rates under abliteration: refusal rates drop by at most 10%, compared to 70-80% drops in baseline models. Comprehensive evaluations of safety and utility demonstrate that extended-refusal fine-tuning effectively neutralizes abliteration attacks while preserving general model performance and enhancing robustness across multiple alignment scenarios.

Harethah Abu Shairah, Hasan Abed Al Kader Hammoud, George Turkiyyah et al.

Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model's safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates - as evaluated by Llama Guard 3 - while preserving the utility of the model on standard benchmarks such as MMLU, HellaSwag, and Arc. Furthermore, we show that ROSI can also re-align 'uncensored' models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results suggest that targeted, interpretable weight steering is a cheap and potent mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.

Harethah Abu Shairah, Somayah AlHarbi, Abdulaziz AlHussein et al.

We introduce ALARB, a dataset and suite of tasks designed to evaluate the reasoning capabilities of large language models (LLMs) within the Arabic legal domain. While existing Arabic benchmarks cover some knowledge-intensive tasks such as retrieval and understanding, substantial datasets focusing specifically on multistep reasoning for Arabic LLMs, especially in open-ended contexts, are lacking. The dataset comprises over 13K commercial court cases from Saudi Arabia, with each case including the facts presented, the reasoning of the court, the verdict, as well as the cited clauses extracted from the regulatory documents. We define a set of challenging tasks leveraging this dataset and reflecting the complexity of real-world legal reasoning, including verdict prediction, completion of reasoning chains in multistep legal arguments, and identification of relevant regulations based on case facts. We benchmark a representative selection of current open and closed Arabic LLMs on these tasks and demonstrate the dataset's utility for instruction tuning. Notably, we show that instruction-tuning a modest 12B parameter model using ALARB significantly enhances its performance in verdict prediction and Arabic verdict generation, reaching a level comparable to that of GPT-4o.

Simon Ghyselincks, Valeriia Okhmak, Stefano Zampini et al.

Reconstructing the structural geology and mineral composition of the first few kilometers of the Earth's subsurface from sparse or indirect surface observations remains a long-standing challenge with critical applications in mineral exploration, geohazard assessment, and geotechnical engineering. This inherently ill-posed problem is often addressed by classical geophysical inversion methods, which typically yield a single maximum-likelihood model that fails to capture the full range of plausible geology. The adoption of modern deep learning methods has been limited by the lack of large 3D training datasets. We address this gap with \textit{StructuralGeo}, a geological simulation engine that mimics eons of tectonic, magmatic, and sedimentary processes to generate a virtually limitless supply of realistic synthetic 3D lithological models. Using this dataset, we train both unconditional and conditional generative flow-matching models with a 3D attention U-net architecture. The resulting foundation model can reconstruct multiple plausible 3D scenarios from surface topography and sparse borehole data, depicting structures such as layers, faults, folds, and dikes. By sampling many reconstructions from the same observations, we introduce a probabilistic framework for estimating the size and extent of subsurface features. While the realism of the output is bounded by the fidelity of the training data to true geology, this combination of simulation and generative AI functions offers a flexible prior for probabilistic modeling, regional fine-tuning, and use as an AI-based regularizer in traditional geophysical inversion workflows.