Are you looking for an exciting opportunity to conduct your thesis with us? On this page, you will find a selection of currently available bachelor thesis topics across our research areas. Explore the topics to find one that matches your interests and background, then follow the application instructions to get in touch with us.

Please note that some of our thesis topics go beyond the foundational knowledge typically covered in a bachelor’s program. These topics offer exciting opportunities to engage with cutting-edge research and explore the state of the art in the field. At the same time, they require curiosity, initiative, and a willingness to independently dive into unfamiliar concepts and invest time in developing a deeper understanding. We encourage students who are open to these challenges and eager to expand their knowledge beyond the standard curriculum to apply.

We look forward to hearing from motivated students interested in contributing to our research!

Bachelor Thesis Topic 1

Benchmarking Decision Models for Intent Detection in Conversational Recommender Systems

Motivation

Conversational recommender systems (CRS) have to understand what the user wants at every turn of a dialogue. Is the user stating a preference, asking for a recommendation, comparing items, or rejecting a suggestion? Today this intent detection step usually is performed by a simple classification model or runs on large language models, which are accurate but slow and expensive at scale. Since recently the class of decision models has emerged, including TypeSafe AI’s Jev, Cloudflare’s Clef and Amazon’s Strands Decider. Instead of generating text, they select from a fixed set of answers in a single pass on par with LLM-level quality, but at a fraction of the cost and latency. Whether these claims hold for the task of intent detection in CRS has not been independently tested.

Description

The thesis carries out a comparative study of decision models on user-intent detection in CRS dialogues. It identifies suitable intent-annotated datasets, such as ReDial/IARD, INSPIRED and MG-ShopDial, and converts them into a common format. It then evaluates open-weight and hosted decision models against established baselines.

Goal

The goal is an open, reproducible benchmarking suite for decision models on intent detection. It should measure both classification accuracy and efficiency (latency, throughput, cost). The outcome shows which models offer the best trade-off between quality and efficiency for intent detection in conversational recommendation.

Bachelor Thesis Topic 2

Smart Product Understanding for Sustainable Fashion Shopping

Motivation

Online fashion shops present product information in very different and often incomplete ways. Material composition, origin, and certifications are scattered across pages or hidden in vague marketing claims like “eco-friendly.” This makes it hard for shoppers, and for recommender systems, to compare products or make informed, sustainable choices.
In this thesis, you will develop a module that automatically extracts structured product information from fashion e-commerce pages and enriches it with a transparent sustainability indicator. You will compare rule-based parsing, structured web data, and LLM-based extraction, and evaluate how reliably key attributes can be identified across different shops.

The module will be potentially integrated into an ongoing PhD project on responsible recommender systems that aim to promote sustainable consumption in online fashion retail.

Task

• Define a product schema (e.g., category, materials, price, brand, certifications)
• Implement and compare extraction approaches for several online shops
• Create a small annotated dataset and evaluate extraction quality
• Design a simple, explainable sustainability score that communicates uncertainty and flags vague green claims

Requirements

• Good programming skills (Python and/or JavaScript)
• Interest in NLP, LLMs, or web technologies
• Interest in sustainability and responsible technology design

Bachelor Thesis Topic 3

Reconstructing a FAIR demolinguistic dataset using population data

Linguistics has seen a rise in the use of computational methods as a result of the development of findable, accessible, interoperable, and reusable (FAIR) datasets, connected through the Cross-Linguistic Linked Data project (CLLD). While the CLLD connects multiple important datasets for describing the world’s linguistic diversity heritage, such as lexicon, phoneme inventories, and phylogenies, there is no FAIR dataset containing language-speaker distributions in the CLLD ecosystem, limiting computational research on linguistic diversity.

This thesis explores the feasibility of reconstructing first-language (L1) speaker distributions by combining demolinguistic data from GlobLingDiv with country-of-birth population data from Eurostat using a weighted distribution approach. The reconstructed L1-distributions will then be evaluated against survey data from the Eurobarometer study Europeans and Their Languages using statistical measures of distributional similarity and linguistic diversity.

Finally, the dataset will be structured according to the Cross-Linguistic Data Formats (CLDF) to enhance interoperability with other CLLD datasets.

Bachelor Thesis Topic 4

A Reusable Evaluation Method for Agentic AI Systems

Motivation

Agentic AI systems combine language models, tool calls and other components to complete tasks. Evaluating their final answers alone gives limited insight into how well the overall system works. A reproducible evaluation method should also consider intermediate actions and support different assessment criteria.

Description

This thesis investigates how existing tools and frameworks can be combined into a reusable evaluation method. It builds on previous master’s theses in our research group and a review of current approaches in research and industry.

You will select suitable tools and develop a prototype workflow, using Python to connect components or add functionality where needed. TUna, TU Wien’s assistant for study-related questions, will serve as the concrete integration and evaluation target. TUna provides access to its live system, Langfuse logs and execution traces, and a ground-truth dataset currently under development. The thesis will use this dataset for evaluation and extend it where needed, for example to assess tool calls.

The workflow should allow interchangeable metrics. Example evaluators include LLM-as-a-judge, decision models for predefined assessment categories, and deterministic NLP checks, such as matching expected entities in reference answers.

Tasks

• Review previous work and existing evaluation tools and methods.
• Combine suitable tools into a reproducible workflow, adding Python extensions where needed.
• Implement a small selection of example metrics and evaluate them using the provided ground-truth dataset, extending it where necessary.
• Integrate and test the prototype with TUna, assessing its limitations and potential reuse in other applications.

Goal

A documented evaluation method and working prototype that demonstrate how existing tools can support systematic evaluation of agentic systems, with flexibility to use different metrics.

Requirements

• Good programming skills in Python
• Basic knowledge of LLMs or NLP
• Interest in agentic systems, software integration and empirical evaluation

How to Apply: 

All the topics are offered within the Christian Doppler Laboratory for Recommender Systems at TU Wien and are available to students enrolled at the TU Wien Faculty of Informatics. If you are interested, please email us at recsys-lab[ät]ec.tuwien.ac.at and include the following documents:

• a short motivational letter (max. 1 page) indicating the preferred topic
• a curriculum vitae (CV)
• a transcript of grades