Talks #15: GHOST DAY Pre-meeting [in English]
Schedule
Tue Apr 14 2026 at 06:00 pm to 09:00 pm
Location
ulica Wierzbięcice 1b, 61-569 Poznań, Polska | Poznan, WP
![Talks #15: GHOST DAY Pre-meeting [in English]](https://cdn.happeningnext.com/events8/banners/2e41191b730dad745a2c35fb4935a987cae989892804a514948eb6cfb7e72a93-rimg-w960-h540-dcffffff-gmir?v=1775765554)
Advertisement
Agenda? Karol Cyganik: Image and Video Quality Metrics ??
? Szymon Czajkowski, Joanna Szczublińska: Large Language Models' Self-Collaboration in the Light of Group Roles Theory ??
? Wiktor Kamzela: Vocabulary-constrained multilingual story generation for language learning ??
? Quiz with prizes - GHOST DAY tickets
? Participant announcements (during quiz sign-ups)
? Pizza & networking
? Location: Allegro, Poznań, Wierzbięcice 1b
? Partner: Allegro
—————————————————————————————
Karol Cyganik: Can you spot a difference? Measuring image and video quality. ??
Abstract:
In the world of machine learning, optimization for L1 or L2 loss functions is a common practice. But does a smaller difference in pixels always mean a better, more realistic image to the human eye? During this presentation, we will examine the evolution of visual quality assessment for images — from classic, analytical metrics to advanced machine learning-based models. We’ll understand why traditional approaches often favor unnaturally blurry graphics and how neural networks learn to evaluate images in a way similar to our visual system. Next, we’ll move on to the practical application of these metrics — specifically, neural rendering methods. When we set an image in motion, the challenges increase dramatically. We’ll see how modern models generate hard-to-detect artifacts that existing evaluation methods struggle to handle.
Bio:
Karol Cyganik obtained his engineering degree in AI from the Poznan University of Technology and a Master's degree in Computer Science from Adam Mickiewicz University in Poznan, Poland. He is currently involved in machine learning research at AMD, where he works as a Senior Software Development Engineer. His research interests include machine learning enabled computer graphics, image reconstruction techniques, and computer vision in general. His research has been productized as AMD Fidelity Super Resolution 4+ and PlayStation Super Resolution (PSSR).
—————————————————————————————
Szymon Czajkowski, Joanna Szczublińska: Beyond the Lone Agent: Evaluating Human-Like Team Roles in LLM Self-Collaboration ??
Abstract:
This presentation explores whether Large Language Models (LLMs) can enhance their problem-solving performance by simulating human-like teamwork dynamics. Using the Calendar Scheduling task from the Natural Plan dataset, we evaluate five distinct system architectures ranging from single-agent setups to multi-agent teams assigned specialized psychological and functional roles
Bio:
Szymon is a graduate of the Poznan University of Technology and currently a Junior Data Science Consultant at KPMG Poland specializing in agentic AI and Data Science solutions. He is a co-creator of the open-source LLaMa-herd interface and gained international expertise in XAI and LLMs at Sapienza University and Bielefeld University. This year he is also the project leader of the GHOST Day conference.
Joanna is a Master’s student in Artificial Intelligence at Poznań University of Technology and AI developer at Orange Polska specializing in cloud-integrated generative models and agentic workflows to enhance business efficiency. Along the way, she won the Hi-Tech Girls 2025 Cloud competition and shared her knowledge at last Women In Tech Camp, supporting others in developing their tech skills.
—————————————————————————————
Wiktor Kamzela: Vocabulary-constrained multilingual story generation for language learning ??
Abstract:
Does language learning have to be boring and rely on repetitive memorization through flashcards? In this talk, we will present a system that uses LLMs to turn vocabulary lists into engaging, personalized stories tailored to the learner’s level. We will walk through the process of constrained text generation, exploring different text rewriting strategies across three languages and four LLMs. We will also present the automated evaluation used to measure the performance of these methods. Our findings show that this approach outperforms the standard constrained beam search, producing stories that are more grammatical, coherent and interesting to read.
Bio:
Wiktor is a Master's student in Artificial Intelligence at Poznań University of Technology and a Junior Machine Learning Researcher at the Poznań Supercomputing and Networking Center (PSNC). His research has resulted in publications at the EMNLP (CORE A*) and EACL (CORE A) conferences, as well as oral and poster presentations at international workshops and conferences. He is an active member of the GHOST student research group, the originator and main coordinator of GHOST Talks, the speaker section leader for the GHOST Day conference, and a co-organizer of the GHOST Academy, an AI course for high school students.
Advertisement
Where is it happening?
ulica Wierzbięcice 1b, 61-569 Poznań, Polska, ulica Wierzbięcice 1, 61-569 Poznań, Polska, Poznan, PolandEvent Location & Nearby Stays:
Know what’s Happening Next — before everyone else does.