Mohamed Ahmed

David Ifeoluwa Adelani, Jesujoba Oluwadara Alabi, Angela Fan et al. · deepmind, mila

Recent advances in the pre-training of language models leverage large-scale datasets to create multilingual models. However, low-resource languages are mostly left out in these datasets. This is primarily because many widely spoken languages are not well represented on the web and therefore excluded from the large-scale crawls used to create datasets. Furthermore, downstream users of these models are restricted to the selection of languages originally chosen for pre-training. This work investigates how to optimally leverage existing pre-trained models to create low-resource translation systems for 16 African languages. We focus on two questions: 1) How can pre-trained models be used for languages not included in the initial pre-training? and 2) How can the resulting translation models effectively transfer to new domains? To answer these questions, we create a new African news corpus covering 16 languages, of which eight languages are not part of any existing evaluation dataset. We demonstrate that the most effective strategy for transferring both to additional languages and to additional domains is to fine-tune large pre-trained models on small quantities of high-quality translation data.

Rishav Hada, Varun Gumma, Adrian de Wynter et al. · microsoft-research

Large Language Models (LLMs) excel in various Natural Language Processing (NLP) tasks, yet their evaluation, particularly in languages beyond the top $20$, remains inadequate due to existing benchmarks and metrics limitations. Employing LLMs as evaluators to rank or score other models' outputs emerges as a viable solution, addressing the constraints tied to human annotators and established benchmarks. In this study, we explore the potential of LLM-based evaluators, specifically GPT-4 in enhancing multilingual evaluation by calibrating them against $20$K human judgments across three text-generation tasks, five metrics, and eight languages. Our analysis reveals a bias in GPT4-based evaluators towards higher scores, underscoring the necessity of calibration with native speaker judgments, especially in low-resource and non-Latin script languages, to ensure accurate evaluation of LLM performance across diverse languages.

Rishav Hada, Varun Gumma, Mohamed Ahmed et al. · microsoft-research

With the rising human-like precision of Large Language Models (LLMs) in numerous tasks, their utilization in a variety of real-world applications is becoming more prevalent. Several studies have shown that LLMs excel on many standard NLP benchmarks. However, it is challenging to evaluate LLMs due to test dataset contamination and the limitations of traditional metrics. Since human evaluations are difficult to collect, there is a growing interest in the community to use LLMs themselves as reference-free evaluators for subjective metrics. However, past work has shown that LLM-based evaluators can exhibit bias and have poor alignment with human judgments. In this study, we propose a framework for an end-to-end assessment of LLMs as evaluators in multilingual scenarios. We create a carefully curated dataset, covering 10 languages containing native speaker judgments for the task of summarization. This dataset is created specifically to evaluate LLM-based evaluators, which we refer to as meta-evaluation (METAL). We compare the performance of LLM-based evaluators created using GPT-3.5-Turbo, GPT-4, and PaLM2. Our results indicate that LLM-based evaluators based on GPT-4 perform the best across languages, while GPT-3.5-Turbo performs poorly. Additionally, we perform an analysis of the reasoning provided by LLM-based evaluators and find that it often does not match the reasoning provided by human judges.

Mohamed Ahmed, Mohamed Abdelmouty, Mingyu Kim et al.

The advancement of Pre-Trained Language Models (PTLMs) and Large Language Models (LLMs) has led to their widespread adoption across diverse applications. Despite their success, these models remain vulnerable to attacks that exploit their inherent weaknesses to bypass safety measures. Two primary inference-phase threats are token-level and prompt-level jailbreaks. Token-level attacks embed adversarial sequences that transfer well to black-box models like GPT but leave detectable patterns and rely on gradient-based token optimization, whereas prompt-level attacks use semantically structured inputs to elicit harmful responses yet depend on iterative feedback that can be unreliable. To address the complementary limitations of these methods, we propose two hybrid approaches that integrate token- and prompt-level techniques to enhance jailbreak effectiveness across diverse PTLMs. GCG + PAIR and the newly explored GCG + WordGame hybrids were evaluated across multiple Vicuna and Llama models. GCG + PAIR consistently raised attack-success rates over its constituent techniques on undefended models; for instance, on Llama-3, its Attack Success Rate (ASR) reached 91.6%, a substantial increase from PAIR's 58.4% baseline. Meanwhile, GCG + WordGame matched the raw performance of WordGame maintaining a high ASR of over 80% even under stricter evaluators like Mistral-Sorry-Bench. Crucially, both hybrids retained transferability and reliably pierced advanced defenses such as Gradient Cuff and JBShield, which fully blocked single-mode attacks. These findings expose previously unreported vulnerabilities in current safety stacks, highlight trade-offs between raw success and defensive robustness, and underscore the need for holistic safeguards against adaptive adversaries.

Bokan Chen, Raiden Hasegawa, Adriaan Hilbers et al.

Shaping multi-megawatt loads, such as data centers, impacts generator dispatch on the electric grid, which in turn affects system CO2 emissions and energy cost. Substantiating the effectiveness of prevalent load shaping strategies, such as those based on grid-level average carbon intensity, locational marginal price, or marginal emissions, is challenging due to the lack of detailed counterfactual data required for accurate attribution. This study uses a series of calibrated granular ERCOT day-ahead direct current optimal power flow (DC-OPF) simulations for counterfactual analysis of a broad set of load shaping strategies on grid CO2 emissions and cost of electricity. In terms of annual grid level CO2 emissions reductions, LMP-based shaping outperforms other common strategies, but can be significantly improved upon. Examining the performance of practicable strategies under different grid conditions motivates a more effective load shaping approach: one that "cherry-picks" a daily strategy based on observable grid signals and historical data. The cherry-picking approach to power load shaping is applicable to any large flexible consumer on the electricity grid, such as data centers, distributed energy resources and Virtual Power Plants (VPPs).

Benedek Fabian, Thomas Edlich, Héléna Gaspar et al.

We apply a Transformer architecture, specifically BERT, to learn flexible and high quality molecular representations for drug discovery problems. We study the impact of using different combinations of self-supervised tasks for pre-training, and present our results for the established Virtual Screening and QSAR benchmarks. We show that: i) The selection of appropriate self-supervised task(s) for pre-training has a significant impact on performance in subsequent downstream tasks such as Virtual Screening. ii) Using auxiliary tasks with more domain relevance for Chemistry, such as learning to predict calculated molecular properties, increases the fidelity of our learnt representations. iii) Finally, we show that molecular representations learnt by our model `MolBert' improve upon the current state of the art on the benchmark datasets.

Rim Assouel, Mohamed Ahmed, Marwin H Segler et al.

Generating novel molecules with optimal properties is a crucial step in many industries such as drug discovery. Recently, deep generative models have shown a promising way of performing de-novo molecular design. Although graph generative models are currently available they either have a graph size dependency in their number of parameters, limiting their use to only very small graphs or are formulated as a sequence of discrete actions needed to construct a graph, making the output graph non-differentiable w.r.t the model parameters, therefore preventing them to be used in scenarios such as conditional graph generation. In this work we propose a model for conditional graph generation that is computationally efficient and enables direct optimisation of the graph. We demonstrate favourable performance of our model on prototype-based molecular graph conditional generation tasks.

Mathias Niepert, Mohamed Ahmed, Konstantin Kutzkov

Numerous important problems can be framed as learning from graph data. We propose a framework for learning convolutional neural networks for arbitrary graphs. These graphs may be undirected, directed, and with both discrete and continuous node and edge attributes. Analogous to image-based convolutional networks that operate on locally connected regions of the input, we present a general approach to extracting locally connected regions from graphs. Using established benchmark data sets, we demonstrate that the learned feature representations are competitive with state of the art graph kernels and that their computation is highly efficient.