Dániel Horváth, HUN-REN/HUN-REN SZTAKI Title: "The AI Scientist in Action: The MIRAI Experiment on AI-Driven Discovery"
Abstract: Recent advances in large language models and agentic AI systems are reshaping the landscape of scientific research. These “AI Scientists” — autonomous, reasoning systems capable of generating hypotheses, interpreting data, and engaging in scientific dialogue — are emerging as new research collaborators.
The MIRAI (Multi-domain Investigation of Research AI Systems) project aims to systematically evaluate cutting-edge AI Scientist architectures across multiple scientific domains. Our goal is to understand how these systems reason, where they fall short, and how they might augment real-world scientific discovery.
Patrik P. Süli, Óbuda University Title: "Obuda University GenAI: A Hybrid Platform for Academic Research and Agentic Workflows"
Abstract: I will present the newly developed GenAI platform of Obuda University, designed to give access to Large Language Models for students and researchers. The system implements a hybrid infrastructure, orchestrating locally hosted open-weight models (such as Qwen 3 Coder) alongside managed services (OpenAI, Google, Anthropic) under a unified interface.
Beyond standard web chat capabilities, the presentation will highlight the platform's API-first approach. By providing OpenAI-compatible endpoints, the system enables users to seamlessly integrate diverse models into agentic workflows, coding assistants (like Cline), and frameworks like LangChain.
Roland Hollós Title: "Syxplain: Neurosymbolic Equation Explorer"
Abstract: I will introduce Syxplain, an agentic algebraic hypothesis maker designed to bridge the gap between deep learning insights and structured network assumptions in scientific knowledge. The system utilizes a Large Language Model (LLM) as a human-compatible scientific reasoner to interpret descriptive statistics, SHAP-based analyses, and variable scatterplots through vision capabilities.
Using this combined quantitative and visual context, the workflow automatically configures PySR with domain-appropriate operators to explore the complexity-loss Pareto frontier. The presentation will detail a secondary reasoning stage where an LLM explains, evaluates, and compares candidate formulas, assessing not only numerical fit but also conceptual coherence within the scientist's domain assumptions. This framework integrates LLMs and symbolic regression into a lightweight workflow for generating transparent, testable symbolic expressions.
Krisztián Sugár Title:"Building an Autonomous AI Scientist: End-to-End Research Automation in Practice"
Abstract: The emergence of large language models has accelerated the vision of fully autonomous “AI Scientist” systems — agents capable of performing end-to-end scientific workflows. My work focuses on constructing a domain-specific autonomous research agent that performs literature search, idea and hypothesis generation, dataset discovery, experiment execution, evaluation, and scientific writing in a unified loop.
I present two practical instantiations of this framework. The first is an AI-driven research idea generator that explores unexplored research directions. The second is a meta-analysis automation pipeline that retrieves evidence, extracts structured data, evaluates study quality, and drafts publication-ready summaries. Both systems are under active development, and each is planned to be published as an independent scientific paper.
