We investigate the generalization boundaries of current Large Multimodal Models (LMMs) under out-of-distribution scenarios and domain-specific tasks. We evaluate their zero-shot generalization across synthetic images, real-world distributional shifts, and specialized datasets like medical and molecular imagery. Empirical results indicate that LMMs struggle with generalization beyond common training domains, limiting their direct application without adaptation. To understand the cause of unreliable performance, we analyze three hypotheses: semantic misinterpretation, visual feature extraction insufficiency, and mapping deficiency. Results identify mapping deficiency as the primary hurdle. To address this problem, we show that in-context learning (ICL) can significantly enhance LMMs’ generalization. We further explore the robustness of ICL under distribution shifts and show its vulnerability to domain shifts, label shifts, and spurious correlation shifts between in-context examples and test data, opening new avenues for overcoming generalization barriers.

A k-testable machine is a finite automaton which recognizes a language L by only seeing a window of size k of each string in L. In this paper we use k-testable machines to recognize probabilistic languages and propose a novel algorithm to learn them. We work in the context of passive learning as our algorithm is based on a finite sample of strings belonging to the target language equipped with frequencies. Because our algorithm learns a probabilistic automaton, the resulting language is less sensitive to noise threshold than García's algorithm. When compared with the ALERGIA learning algorithm, our method provides a better result in the case of the target language being a k-testable language. In fact, in this case, for the given window k we can learn at the limit the target language exactly.

The European Telecommunications Standards Institute (ETSI) has published a Group Specification document outlining Network Function Virtualization Management and Orchestration (NFV-MANO) systems, advocating an open environment for the purpose of integrating multiple vendor solutions into one NFV ecosystem. To align with this initiative, there is a need to transform the specification into real products with the necessary components and interfaces following the ETSI reference architecture. To support this effort, in this article we share our experience in the exploration of further refining and developing internal components of the ETSI NFV-MANO functional blocks with identified open integration interfaces. We focus on the detailed analysis of the life cycle management of the Virtual Network Functions (VNFs), which is the most fundamental and important functionality of NFV-MANO. Based on our analysis, we propose an abstract VNF Manager Integration Interface (VNFM Integration Interface) at the southbound of VNFM, to integrate with various clouds and systems to complete the life cycle management of the VNFs. In order to achieve advanced VNF life cycle management such as auto-scaling, a model-based monitoring solution is adopted for monitoring the VNFs across multiple layers of NFV architecture. To serve as a proof of concept, we automatically deploy and scale a sample VNF using the proposed software architecture and interfaces. With the prototyping experience using the VNF model as a plugin, to integrate with traditional virtualization and physical infrastructure hierarchy as a full topology model, we evaluate using a model-based monitoring system Dell Foglight™ for realizing NFV monitoring functionalities as part of NFV-MANO architecture framework.