This paper addresses the problem of automated generation of
structured descriptions of master thesis content using
large language models (LLMs). In academic practice at the
University of Belgrade, School of Electrical Engineering,
this section describing the structure and content of a
thesis is typically written manually in a formal and
standardized style, requiring consistent coverage of all
thesis chapters while preserving clarity and coherence. To
alleviate the burden of doing such labor-intensive work for
the academic staff, we propose a LLM-based approach. We
propose a structure-aware approach that leverages the
inherent organization of the thesis by decomposing the task
into chapter-level description generation, followed by
section assembly and stylistic refinement. Instead of
relying on a single end-to-end generation step, the
proposed pipeline processes each chapter independently,
ensuring better control over content coverage and reducing
the risk of hallucinations and structural inconsistencies.
The approach is evaluated on a collection of 5 master
theses using multiple LLMs, including Qwen, Gemma, and
Mistral. Evaluation combines lexical, semantic, and
composite metrics, as well as LLM-based factual consistency
assessment. Experimental results show that the proposed
structure-aware pipeline produces coherent and semantically
accurate descriptions, with Mistral achieving the highest
scores of the chosen models.

Code comments play an important role in software
development and programming education, as they help explain
implementation intent and may indicate whether students
understand the underlying code. However, automated
assessment of comment correctness becomes more challenging
in bilingual educational settings, where students may write
comments in either English or a local language such as
Serbian. This paper investigates the use of open large
language models for binary classification of code-comment
correctness in English and Serbian. We construct a
bilingual dataset in which the same code fragments are
paired with English and Serbian comments sharing the same
correctness labels, and evaluate several code-oriented and
general-purpose LLMs using a few-shot prompting strategy.
The results show that Qwen3 and Qwen2.5-Coder achieve the
strongest performance across both language conditions,
while other models exhibit larger variation, particularly
on Serbian comments. Pairwise McNemar tests further show
that several observed differences between models are
statistically significant. Overall, the findings suggest
that LLM-based code-comment evaluation is a promising
direction for programming education, but also highlight the
need for multilingual evaluation when applying such systems
in local-language educational contexts

In the delivery of higher education, there is a need for
students to work in real conditions, with real data and
real-world examples. Therefore, it is necessary to
increasingly enrich practical teaching in new study
programs with real-life practice. However, this presents a
challenge, since industrial, business, and other real-world
data, as well as insight into actual processes and
configurations, are in most cases unavailable or restricted
due to the nature of operations. For this reason, this
paper highlights research projects in which higher
education institutions often participate, and which can
compensate for these limitations. Such projects can provide
students, within future study programs, with opportunities
to acquire practical, real-world knowledge that will
genuinely benefit their future engineering work. Through
real-life examples, the paper will demonstrate both the
primary and extended potential of this approach, as well as
its sustainability.

In the modern higher education environment of the field of
Information Technology (IT), the application of modern
technological tools is more and more oriented toward the
integration of the existing gap between theoretical and
applied knowledge. One of the major problems in the
development of IT courses is the process of introductory
programming, often associated with high levels of cognitive
load and varied levels of student performance. This paper
aims to investigate the pedagogical potential and
operational application of the integration of AI-based
tools of virtual agents in the process of programming
education. This is based on the theoretical background of
the constructivist approach and the self-regulated learning
(SRL) model. The empirical part of the research, carried
out among IT students, examines their views concerning the
usefulness of virtual assistants in overcoming syntax
barriers and improving problem-solving fluency. The
research findings suggest that although AI tools greatly
alleviate cognitive load and motivate learners, it is
necessary to apply a didactic approach in order to develop
genuine computational thinking. The research findings
contribute to the discussion aimed at improving the quality
of IT study programs and making them compatible with the
changing needs of the global labor market.

Gradivo predmeta Performanse računarskih sistema obuhvata
niz matematičkih modela čije ponašanje studenti teško stiču
samo kroz klasičnu nastavu. U ovom radu se opisuje veb
aplikacija razvijena u Python/Streamlit okruženju koja
studentima omogućava da eksperimentišu sa parametrima tih
modela i odmah vide rezultat na grafiku. Aplikacija je
podeljena u tri modula, a to su performanse hardvera,
sistemi masovnog opsluživanja i otvorene mreže, i pokriva
celokupan kurs, od Amdahl-ovog zakona do MVA algoritma. Rad
daje pregled srodnih alata (JMT, JMCH, Jupyter notebook
rešenja), opisuje arhitekturu aplikacije, prikazuje pet
odabranih vežbi i razmatra kako se alat može uklopiti u
nastavu.

Ovaj rad predstavlja simulator razvijen kao veb aplikacija,
a namenjen boljem razumevanju algoritama pretrage u
nastavi. Algoritmi pretrage često predstavljaju izazov za
studente zbog visokog nivoa apstraktnosti i složenosti
izvršavanja. Postojeća rešenja su uglavnom ograničena u
pogledu vizualizacije, kontrole simulacije i
interaktivnosti. U radu je predložen simulator koji
omogućava kreiranje grafova, podešavanje težina grana i
heuristike, odabira algoritma i izvršavanje algoritma uz
detaljan prikaz svakog koraka. Predloženi pristup
predstavlja koristan alat u nastavi koji doprinosi dodatnom
razumevanju algoritama pretrage nad grafovima.