For the successful deployment of large language models (LLMs) in business and scientific environments, models require more than just knowledge—they must be aligned with ethical standards and specific user preferences. This practical workshop will cover the fundamentals of alignment, reinforcement learning from human feedback (RLHF), bias mitigation, and adaptation to user needs.

Location: JSI E-room Teslova uica 30, Ljubljana

Time: 10:00 – 14:00

More information and registration…

The discussion focused on replacing anglocentric LLM evaluation benchmarks—which inject foreign values, biases, and cultural contexts—with continuous, language-specific testing. The panelists underscored that evaluation frameworks must be treated as essential linguistic infrastructure, tailored directly to Slovenian cultural and societal values.

Slides are already available.

Read more about the conference.

A significant part of the discussion also focused on the business support ecosystem, highlighting the Slovenian AI Factory (SLAIF) and the Slovenian Competence Center for Artificial Intelligence (KCUI). The role of the KCUI lies precisely in bridging research expertise with the needs of the business sector and in fostering the transfer of best practices and AI solutions into practical application.

📍 𝗕𝗿𝗱𝗼 𝗽𝗿𝗶 𝗞𝗿𝗮𝗻𝗷𝘂 📅 𝟭𝟭 𝗦𝗲𝗽𝘁𝗲𝗺𝗯𝗲𝗿 𝟮𝟬𝟮𝟲

In addition to the formal sessions, the meeting featured open technical discussions that fostered collaboration across disciplines and strengthened alignment among consortium members.

Beyond the technical and scientific progress, the General Assembly highlighted the strong commitment and collaborative spirit of the consortium. The event provided a valuable opportunity to reconnect in person, consolidate ongoing work, and discuss future opportunities for extending the impact of TRUSTroke beyond the current project framework.

You can find out more about the meeting at the link: https://mailchi.mp/e5d7080ae29f/6th-trustroke-nl

The authors, including Marko Debeljak and Vid Podpečan from the EcoEnvAI research group, from E8 at the Jožef Stefan Institute, alongside an international team of collaborators, developed an LLM-assisted workflow to automatically extract and structure knowledge from scientific abstracts, enabling rapid evidence mapping in soil science.

The study demonstrates the application of AI in supporting environmental research, identifying knowledge gaps, and accelerating scientific synthesis.

The open-access paper is available at: https://www.sciencedirect.com/science/article/pii/S0016706126002181

This framework improves off-the-shelf LLM embeddings (such as OpenAI text-embedding) for classification tasks by incorporating global knowledge from Wikidata KGs and local knowledge learned from task-specific knowledge graphs. High-dimensional fusion can be matched by searching for low-dimensional weighted modality contributions via Bayesian Optimization. Furthermore, tabular foundation models serve as effective surrogates, achieving results comparable to more expensive AutoML runs at a fraction of the computational cost.

📄 Link: https://lnkd.in/dUmENV-e

The article covers the AIDA system (Agricultural Intelligence and Decision Assistant), developed to assist farmers and agricultural advisors with expert information, regulatory compliance, and decision-making. Designed specifically for the Slovenian language and agricultural context, the platform integrates local knowledge sources.

The project was highlighted by journalist Jana Petkovšek Štakul for Agrobiznis and is being developed by the EcoEnvAI group in collaboration with ITC Murska Sobota and the KGZS Murska Sobota Institute.

Read more…

This new release focuses heavily on reproducibility and ease of use. Key updates include:
– Easier Sharing: Seamlessly share your code, configuration, data, and experimental setups.
– Grammars and constraints: Add custom constraints to your search space effortlessly.
– Performance Boosts: Significant speed and efficiency gains.

Check out the package and documentation here: https://lnkd.in/dw59pkbi