Model evolution enables learning from feedback to refine experiences and update skills, transforming models from having no domain knowledge to becoming domain experts. However, there is currently no unified and effective method for guiding this evolutionary process. To address this gap, we propose the Meteor method, which includes three training phases: weak-to-strong data distillation, iterative training, and self-evolution strategies. Each phase maximizes the model's inherent domain capabilities, allowing it to autonomously refine its domain knowledge and enhance performance. Experiments demonstrate that our approach significantly improves accuracy, completeness, relevance, coherence, and reliability across domain-specific tasks.

Graph Neural Networks (GNNs) have emerged as a powerful tool to capture intricate network patterns, achieving success across different domains. However, existing GNNs require careful domain-specific architecture designs and training from scratch on each dataset, leading to an expertise-intensive process with difficulty in generalizing across graphs from different domains. Therefore, it can be hard for practitioners to infer which GNN model can generalize well to graphs from their domains. To address this challenge, we propose a novel cross-domain pretraining framework, "one model for one graph," which overcomes the limitations of previous approaches that failed to use a single GNN to capture diverse graph patterns across domains with significant gaps. Specifically, we pretrain a bank of expert models, with each one corresponding to a specific dataset. When inferring to a new graph, gating functions choose a subset of experts to effectively integrate prior model knowledge while avoiding negative transfer. Extensive experiments consistently demonstrate the superiority of our proposed method on both link prediction and node classification tasks.

This technical report presents the details of the development and evaluation of the Sabiá-3 and Sabiazinho-3 models. We trained them on a large corpus of documents written in Portuguese, with a special focus on Brazil-related resources. Through training, models were exposed to information relevant to Brazilian culture, history, and context. The main objective was to have a specialized model that is aware of the linguistic nuances, societal norms, and regional variations unique to the country. Throughout this report, we show that this specialization allows the models to perform better in knowledge-intensive tasks. We applied an approach of continual learning by leveraging a "generalist" model that already acquired some level of language understanding and reasoning abilities, and then further trained it on our corpus of high-quality data relevant to the Brazilian context. The development consisted of two main phases: (1) the pre-training phase, in which we further train a pre-trained model on specialized data following a self-supervised learning strategy optimizing for the next token prediction objective, and (2) the post-training phase where the model is tuned to follow instructions and align to human preferences. Compared to our previous release, Sabiá-2 [5], we have collected a significantly larger volume of data for pre-training.

Natural language processing skills of Large Language Models (LLMs) are unprecedented, having wide diffusion and application in different tasks. This pilot study focuses on specializing ChatGPT behavior through a Retrieval-Augmented Generation (RAG) system using the OpenAI custom GPTs feature. The purpose of our chatbot, called Unimib Assistant, is to provide information and solutions to the specific needs of University of Milano-Bicocca (Unimib) students through a question-answering approach. We provided the system with a prompt highlighting its specific purpose and behavior, as well as university-related documents and links obtained from an initial need-finding phase, interviewing six students. After a preliminary customization phase, a qualitative usability test was conducted with six other students to identify the strengths and weaknesses of the chatbot, with the goal of improving it in a subsequent redesign phase. While the chatbot was appreciated for its user-friendly experience, perceived general reliability, well-structured responses, and conversational tone, several significant technical and functional limitations emerged. In particular, the satisfaction and overall experience of the users was impaired by the system's inability to always provide fully accurate information. Moreover, it would often neglect to report relevant information even if present in the materials uploaded and prompt given. Furthermore, it sometimes generated unclickable links, undermining its trustworthiness, since providing the source of information was an important aspect for our users. Further in-depth studies and feedback from other users as well as implementation iterations are planned to refine our Unimib Assistant.

Function as a Service (FaaS) is poised to become the foundation of the next generation of cloud systems due to its inherent advantages in scalability, cost-efficiency, and ease of use. However, challenges such as the need for specialized knowledge and difficulties in building function workflows persist for cloud-native application developers. To overcome these challenges and mitigate the burden of developing FaaS-based applications, in this paper, we propose a mechanism called Action Engine, that makes use of Tool-Augmented Large Language Models (LLMs) at its kernel to interpret human language queries and automates FaaS workflow generation, thereby, reducing the need for specialized expertise and manual design. Action Engine includes modules to identify relevant functions from the FaaS repository and seamlessly manage the data dependency between them, ensuring that the developer's query is processed and resolved. Beyond that, Action Engine can execute the generated workflow by feeding the user-provided parameters. Our evaluations show that Action Engine can generate workflows with up to 20\% higher correctness without developer involvement. We notice that Action Engine can unlock FaaS workflow generation for non-cloud-savvy developers and expedite the development cycles of cloud-native applications.

Qualitative Spatial Reasoning is a well explored area of Knowledge Representation and Reasoning and has multiple applications ranging from Geographical Information Systems to Robotics and Computer Vision. Recently, many claims have been made for the reasoning capabilities of Large Language Models (LLMs). Here, we investigate the extent to which a set of representative LLMs can perform classical qualitative spatial reasoning tasks on the mereotopological Region Connection Calculus, RCC-8. We conduct three pairs of experiments (reconstruction of composition tables, alignment to human composition preferences, conceptual neighbourhood reconstruction) using state-of-the-art LLMs; in each pair one experiment uses eponymous relations and one, anonymous relations (to test the extent to which the LLM relies on knowledge about the relation names obtained during training). All instances are repeated 30 times to measure the stochasticity of the LLMs.

Vision-Language Models (VLMs) have great potential in medical tasks, like Visual Question Answering (VQA), where they could act as interactive assistants for both patients and clinicians. Yet their robustness to distribution shifts on unseen data remains a critical concern for safe deployment. Evaluating such robustness requires a controlled experimental setup that allows for systematic insights into the model's behavior. However, we demonstrate that current setups fail to offer sufficiently thorough evaluations, limiting their ability to accurately assess model robustness. To address this gap, our work introduces a novel framework, called SURE-VQA, centered around three key requirements to overcome the current pitfalls and systematically analyze the robustness of VLMs: 1) Since robustness on synthetic shifts does not necessarily translate to real-world shifts, robustness should be measured on real-world shifts that are inherent to the VQA data; 2) Traditional token-matching metrics often fail to capture underlying semantics, necessitating the use of large language models (LLMs) for more accurate semantic evaluation; 3) Model performance often lacks interpretability due to missing sanity baselines, thus meaningful baselines should be reported that allow assessing the multimodal impact on the VLM. To demonstrate the relevance of this framework, we conduct a study on the robustness of various fine-tuning methods across three medical datasets with four different types of distribution shifts. Our study reveals several important findings: 1) Sanity baselines that do not utilize image data can perform surprisingly well; 2) We confirm LoRA as the best-performing PEFT method; 3) No PEFT method consistently outperforms others in terms of robustness to shifts. Code is provided at https://github.com/IML-DKFZ/sure-vqa.

Recent years have seen a significant surge in complex AI systems for competitive programming, capable of performing at admirable levels against human competitors. While steady progress has been made, the highest percentiles still remain out of reach for these methods on standard competition platforms such as Codeforces. Here we instead focus on combinatorial competitive programming, where the target is to find as-good-as-possible solutions to otherwise computationally intractable problems, over specific given inputs. We hypothesise that this scenario offers a unique testbed for human-AI synergy, as human programmers can write a backbone of a heuristic solution, after which AI can be used to optimise the scoring function used by the heuristic. We deploy our approach on previous iterations of Hash Code, a global team programming competition inspired by NP-hard software engineering problems at Google, and we leverage FunSearch to evolve our scoring functions. Our evolved solutions significantly improve the attained scores from their baseline, successfully breaking into the top percentile on all previous Hash Code online qualification rounds, and outperforming the top human teams on several. Our method is also performant on an optimisation problem that featured in a recent held-out AtCoder contest.

The effectiveness of Large Language Models (LLMs) significantly relies on the quality of the prompts they receive. However, even when processing identical prompts, LLMs can yield varying outcomes due to differences in their training processes. To leverage the collective intelligence of multiple LLMs and enhance their performance, this study investigates three majority voting strategies for text classification, focusing on phishing URL detection. The strategies are: 1) a prompt-based ensemble, which utilizes majority voting across the responses generated by a single LLM to various prompts 2) a model-based ensemble, which entails aggregating responses from multiple LLMs to a single prompt; and 3) a hybrid ensemble, which combines the two methods by sending different prompts to multiple LLMs and then aggregating their responses. Our analysis shows that ensemble strategies are most suited in the cases where individual components--whether prompts or LLMs--exhibit equivalent performance levels. However, when there is a significant discrepancy in individual performance, the effectiveness of the ensemble method may not exceed that of the highest-performing single LLM or prompt. In such instances, opting for ensemble techniques is not recommended.

Video Large Language Models (Video-LLMs) have recently shown strong performance in basic video understanding tasks, such as captioning and coarse-grained question answering, but struggle with compositional reasoning that requires multi-step spatio-temporal inference across object relations, interactions, and events. The hurdles to enhancing this capability include extensive manual labor, the lack of spatio-temporal compositionality in existing data and the absence of explicit reasoning supervision. In this paper, we propose STEP, a novel graph-guided self-training method that enables Video-LLMs to generate reasoning-rich fine-tuning data from any raw videos to improve itself. Specifically, we first induce Spatio-Temporal Scene Graph (STSG) representation of diverse videos to capture fine-grained, multi-granular video semantics. Then, the STSGs guide the derivation of multi-step reasoning Question-Answer (QA) data with Chain-of-Thought (CoT) rationales. Both answers and rationales are integrated as training objective, aiming to enhance model's reasoning abilities by supervision over explicit reasoning steps. Experimental results demonstrate the effectiveness of STEP across models of varying scales, with a significant 21.3\% improvement in tasks requiring three or more reasoning steps. Furthermore, it achieves superior performance with a minimal amount of self-generated rationale-enriched training samples in both compositional reasoning and comprehensive understanding benchmarks, highlighting the broad applicability and vast potential.