We understand little about how capabilities acquired in one language carry over to another, or what governs this transfer: evaluations rely on incomparable, saturation-prone datasets and rarely examine its determinants jointly. Identifying what predicts transfer would let us avoid exhaustive evaluation across all language pairs and let developers target the factors that limit performance in low-resource languages. To evaluate cross-lingual capability transfer, we introduce Multilingual GSM-Symbolic, an extensible multilingual mathematical dataset covering 30,000 item-matched question-answer pairs and spanning 15 languages. It utilises symbolic templates to prevent overfitting and ensure generalisation by allowing generation of millions of high-quality variations from a single sample. Using Multilingual GSM-Symbolic, we quantify the largest determinants of capability as model size (β= 1.77), language resource level (β= 0.77), reasoning (β= 0.67) and typological distance (β= -0.25). This joint estimation allows these determinants to be expressed in terms of one another: a 32B model evaluated in Marathi performs like a 10B model in English. Our findings have important implications for model developers, showing that model size and reasoning narrow the performance gap between low- and high-resource languages (β= -0.27 and β= -0.20, respectively), while similar levers have little or no effect on typologically distant languages. Overall, our analysis framework explains 92
The rapid advancements in AI and Machine Learning necessitate a robust computational infrastructure to support cutting-edge research and industrial applications. From the academic and industrial AI community perspective, voiced in the recent ELISE project, the European AI platform is recommended to center around the EuroHPC growing ecosystem. It should be user-driven, easily accessible, powerful, and compliant with European regulations. AI-optimized and dedicated supercomputers for the European AI community are also coming, in addition to upgrading partitions of existing EuroHPC systems to 'AI enabled' stage. Related calls have been initiated in September 2024. Further, conventional EuroHPC systems are suggested to be extended with quantum computing, edge AI, and neuromorphic computing to cater to AI models deployed on network edge devices and sustainability in the long run. The challenges are presented in three case studies, ranging from training Transformers on HPC to LLMs trained federally across three different Euro HPC systems to recent results on hybrid classical-quantum application. This paper concludes with case studies results-informed next steps believed to benefit AI practitioners and the broader AI community.
We introduce graphical complexes of groups, which can be thought of as a generalisation of Coxeter systems with 1-dimensional nerves. We show that these complexes are strictly developable, and we equip the resulting Basic Construction with three structures of non-positive curvature: piecewise linear CAT(0), C(6) graphical small cancellation, and a systolic one. We then use these structures to establish various properties of the fundamental groups of these complexes, such as biautomaticity and Tits Alternative. We isolate an easily checkable condition implying hyperbolicity of the fundamental groups, and we construct some non-hyperbolic examples. We also briefly discuss a parallel theory of C(4)-T(4) graphical complexes of groups and outline their basic properties.
This paper presents the FactoryML framework that simplifies the deployment and integration of Machine Learning (ML) models in manufacturing factory environments. FactoryML facilitates packaging of ML models into a portable format and it facilitates the communication of deployed ML models in factory environments via Programmable Logic Controllers. In general FactoryML reduces the barrier to take learned models from a research and development side into an operational setting. The value of FactoryML has been demonstrated in a case study with a Danish company as well.
Secure Multi-Party Computation (MPC) constructions typically allow computation over a finite field or ring. While useful for many applications, certain real-world applications require the usage of decimal numbers. While it is possible to emulate floating-point operations in MPC, fixed-point computation has gained more traction in the practical space due to its simplicity and efficient realizations. Even so, current protocols for fixed-point MPC still require computing a secure truncation after each multiplication gate. In this paper, we show a new paradigm for realizing fixed-point MPC. Starting from an existing MPC protocol over arbitrary, large, finite fields or rings, we show how to realize MPC over a residue number system (RNS). This allows us to leverage certain mathematical structures to construct a secure algorithm for efficient approximate truncation by a static and public value. We then show how this can be used to realize highly efficient secure fixed-point computation. In contrast to previous approaches, our protocol does not require any multiplications of secret values in the underlying MPC scheme to realize truncation but instead relies on preprocessed pairs of correlated random values, which we show can be constructed very efficiently, when accepting a small amount of leakage and robustness in the strong, covert model. We proceed to implement our protocol, with SPDZ [28] as the underlying MPC protocol, and achieve significantly faster fixed-point multiplication.