University of Primorska
Faculty of Mathematics, Natural Sciences and Information Technologies
- 1st seminar (30 minutes):
PRESENTER: Kaja PRAPROTNIK, a PhD student in Computer Science at UP FAMNIT.
TITLE:
Influencing Sentiment in LLM Output via Token Selection Manipulation (part 2)
ABSTRACT:
We conduct a human evaluation of sentiment steering in Llama-3.1-8B-Instruct, using arena-style pairwise comparison in which evaluators choose which of two outputs from the same prompt feels more positive or negative. The steering biases the logits of tokens drawn from a sentiment lexicon at each decoding step, without additional training, comparing an additive bias with a multiplication and division formulation. Outputs are also scored automatically with the lexicon-based sentiment analysis tool VADER, and analysed with Friedman tests and Wilcoxon signed-rank comparisons against the baseline.
ABOUT THE PRESENTER:
Kaja Praprotnik is a PhD student at UP FAMNIT under the supervision of Prof. Jernej Vičič. Her research focuses on natural language processing with a strong focus on artificial intelligence.
- 2nd seminar (30 minutes):
PRESENTER: Jovan PAVLOVIĆ, PhD student in Computer Science at UP FAMNIT.
TITLE:
A retrieval-augmented generation framework designed to improve evidence retrieval for complex, multi-hop reasoning tasks
ABSTRACT:
This presentation focuses on a retrieval-augmented generation (RAG) framework designed to improve evidence retrieval for complex, multi-hop reasoning tasks. Conventional RAG systems often treat all retrieved passages as equally reliable. GraphRAG approaches augment retrieval with knowledge graphs that represent relationships between entities and support multi-step reasoning, but they may rely on pre-existing, external knowledge graphs and LLM decision-making during the retrieval process. Our method, called SA-RAG, aims to improve knowledge retrieval using a spreading activation algorithm to search for relevant information in an interconnected document corpus, without relying on an LLM to guide the process or on human-crafted knowledge graphs. Experiments on multi-hop question-answering tasks show that SA-RAG performs comparably to or better than state-of-the-art methods and improves answer correctness by approximately 12 percentage points over naive vector-based retrieval.
ABOUT THE PRESENTER:
Jovan Pavlović is a Ph.D. student and teaching assistant at UP FAMNIT, with research interests spanning network science and machine learning, graph-enhanced retrieval-augmented generation, and graph neural networks.
Both seminars will be held in English, starting at 4 PM in the classroom FAMNIT-VP3.
