Postdoctoral Fellowship in Predictive World Models for Image Generation

Is the Job related to staff position within a Research Infrastructure?: Yes

Offer Description

• Research Topic: research on methods, technologies, and practices for self-supervised joint-embedding predictive architectures, predictive world models, and generative image synthesis, including latent-space forecasting and conditioned diffusion for controllable visual generation.
• Planned Fellowship Supervisors: Assoc. Prof. Dr. Valentas Gružauskas and Assoc. Prof. Dr. Linas Petkevičius

The candidate will be a part of a multidisciplinary and collaborative team that can engage with international partners. 

• Monthly Salary: €3,634.00 gross (before tax).
• Duration of Work Contract: 12 months.
• Please note: candidates who are unable to work in Lithuania, at the premises of the institutes of the Faculty of Mathematics and Informatics at Vilnius University, will not be considered.

Prospective candidates interested in collaborating with Assoc. Prof. Valentas Gružauskas and Assoc. Prof. Linas Petkevičius in their research group will be hosted at the AI Methods Lab of the Institute of Computer Science, Faculty of Mathematics and Informatics, and can expect to work on self-supervised representation learning, predictive world models and generative image synthesis. Research spans both fundamental methodological development and practical applications on large-scale image collections, including satellite image time series. 

Assoc. Prof. Dr. Valentas Gružauskas works at the intersection of AI and computational modelling, with a focus on agent-based modelling, socio-economic simulation, and modern AI methods (incl. LLM-driven agentic workflows) for decision support in complex systems. He has extensive teaching and supervision experience in AI-related topics and supervises technical projects involving LLM agents, evaluation and controllability. His research emphasises reproducible computational experiments, sensitivity/uncertainty analysis, and methodological rigour when linking empirical evidence to simulation. The postdoctoral fellow will be guided in a highly technical manner, with a strong emphasis on rigorous experimental design, ablation practice and reproducibility.

Profiles:Linas Petkevičius Institute profile

Assoc. Prof. Linas Petkevičius has a master's degree in statistics and PhD in computer science (2020, Vilnius University) with competencies in both informatics and statistics. After his doctorate studies, Dr. Petkevičius works with machine learning. He is the co-author of 13 articles published in scientific journals with a citation index (IF) in the Web of Science (WoS) database. Previously he was an expert at the Agency for Science and the Austrian Science Foundation (FWF), a board member of the European AI Forum and a member of the editorial board of the journal Nonlinear Analysis: Modeling and Control (Q1). From 2026 he is the Head of Institute of Computer Science at the Faculty of Mathematics and Informatics. Dr. Petkevičius is teaching the Deep learning and introduction to Quantum computing courses. Currently, he is the team lead of national research project: ""New generation multi-task recognition from satellite image algorithms for climate monitoring"" (NEUTRINO).

Key Responsibilities & Research Lines:

The postdoctoral fellow will drive research on predictive world models for image generation — systems that learn to anticipate future visual states in an abstract representation space and then render them as high-fidelity images. The work is methodological in focus, with large-scale image time series, including Earth observation archives, serving as the principal experimental testbed. Over the 12 months of the fellowship, core responsibilities and research directions include:

• Designing, implementing and rigorously evaluating self-supervised joint-embedding predictive architectures that forecast future latent states directly in representation space rather than in pixel space — covering encoder and predictor design, target-encoder and exponential-moving-average stabilisation, and training objectives that prevent representation collapse.
• Coupling learned world models to generative image backbones, in particular latent diffusion and flow-matching models, through lightweight conditioning adapters, cross-attention interfaces and parameter-efficient fine-tuning, so that predicted semantics govern what is generated while a largely frozen backbone supplies photorealistic texture.
• Investigating multimodal conditioning signals for controllable generation — including vision-language scene descriptions, structural priors and learned embeddings — and studying how linguistic and latent-spatial signals can be fused, weighted and disentangled to steer generation without inducing semantic drift or hallucinated structure.
• Addressing long-horizon consistency in autoregressive rollout: characterising and mitigating error accumulation, variance propagation and structural drift when a world model is applied recursively over many prediction steps.
• Advancing evaluation methodology for generative prediction, including the perception–distortion trade-off, structural and perceptual quality measures, and ablation protocols that isolate the contribution of individual architectural components.
• Implementing the resulting methods within reproducible, PyTorch-based experimentation pipelines on HPC infrastructure, supporting large-scale training, hyperparameter optimisation, benchmarking and open-source release.
• Contributing to high-impact scientific publications, presentations at international conferences, and the preparation of competitive research proposals, while actively participating in the collaborative, international environment of the AI Methods Lab.

Mentoring and knowledge transfer activities, including support for student projects and lab seminars in self-supervised learning, world models and generative vision.

Where to apply

E-mail: mokslo.prodekanas@mif.vu.lt

Requirements

Research Field: Computer science » Informatics

Education Level: PhD or equivalent

Skills/Qualifications

• Research Expertise: strong background in deep learning, computer vision or generative modelling.
• Experience with Vision-Language Models (VLMs) and multi-modal AI systems.
• Programming Skills in Python, PyTorch, and modern ML frameworks.
• Experience with self-supervised or representation learning methods, and with training models on large-scale image or video datasets.
• Language Skills: good working knowledge of English (B2 level or higher), with strong written and oral communication skills.
• Familiarity with generative image models — for example latent diffusion, flow matching or autoregressive architectures.
• Publications in machine learning, AI, computer vision, or related areas.
• Ability to design and implement AI research prototypes.
• Strong analytical thinking and problem-solving skills.
• Interest in world models, predictive architectures and controllable visual generation.
• Good communication skills and ability to work in international research teams.
• Motivation to contribute to cutting-edge AI research and scientific innovation.

Desirable Assets (Considered as Advantages):

• Experience in academic mentoring, including leading practical seminars or co-supervising student theses within an international university environment.
• Demonstrated experience or motivation in drafting technical grant proposals for international funding agencies.
• Experience with distributed or multi-GPU training of large vision models on HPC infrastructure.

Specific Requirements

Strong preference will be given to candidates who obtained their PhD degree from prestigious international research institutions outside of Lithuania that possess an established, world-class research framework.

Internal Application form(s) needed

APPLICATION FOR VU POSTDOCTORAL FELLOWSHIPS (FORM).pdf
English
(61.12 KB - PDF)
Download

Additional Information

Benefits

A great opportunity to conduct research in a friendly team, with good working conditions and the chance to gain new professional experience.

Eligibility criteria

Applicants must satisfy the eligibility requirements for postdoctoral fellowships at Vilnius University:

1. The candidate must hold a doctoral degree awarded by a foreign institution (preferred) or a Lithuanian science institution other than Vilnius University.
2. No more than 5 years must have elapsed since the awarding of the PhD degree. This 5-year window can be formally extended for documented periods of maternity, paternity, or parental leave.

Selection process

The evaluation follows a standard, transparent merit-based academic review process. Candidates will be ranked based on the alignment of their competencies with the research description and the quality of their scientific publication record.

Please note: To optimize the administrative process, only shortlisted candidates who pass the initial screening phase will be contacted and invited for an online interview. We kindly thank all applicants for their time and interest in this position.

Additional comments

Required Documents:

Applicants must submit a single application package via email, which consists of the official application form and its mandatory attachments:

1. Official Application Form: The completed and signed Application Form for Vilnius University Postdoctoral Fellowships (ensuring the fields "Planned fellowship supervisor" and "Title of the fellowship" match the details of the Offer).
2. Mandatory Form Attachments (as specified in the application template):
   
   • Curriculum Vitae (CV): A free-form description of your life and scientific activities.
   • List of Publications: A structured list highlighting up to 10 of your most important scientific publications and/or patents.
3. Digital Copy of PhD Diploma: Required by the university administration for the formal 5-year eligibility window verification check.

Application Submission:

Applications must be submitted electronically to the Vice-Dean for Science of the Faculty, via email at: mokslo.prodekanas@mif.vu.lt .

Important: Please ensure the email subject line states exactly "postdoctoral fellowship" to guarantee proper routing, indexing, and official registration.

Work Location(s)

• Number of offers available: 1
• Company/Institute: Institute of Computer Science, Faculty of Mathematics and Informatics of Vilnius University
• Country: Lithuania
• State/Province: Vilnius District
• City: Vilnius
• Postal Code: LT-03225
• Street: Naugarduko str. 24

Contact

• State/Province: Vilnius District
• City: Vilnius
• Website: https://www.vu.lt/en/
• Street: 3 Universiteto street
• Postal Code: LT-01513