Constantinos Karouzos

Hi! I’m Constantinos.

A PhD student at the UKRI CDT in Speech and Language Technologies and their Applications, at the School of Computer Science in the University of Sheffield, UK.

I’m working on preference optimisation and domain adaptation. My supervisor is Prof. Nikos Aletras.

In 2020, I graduated from School of ECE at the National Technical University of Athens, obtaining an Integrated M.Eng. in Electrical and Computer Engineering. My thesis title is “Unsupervised Domain Adaptation for Natural Language Processing” and my advisor was Prof. Alex Potamianos. Previously, I worked as a Machine Learning Engineer at Behavioral Signals, a startup aiming to bridging the communication gap between humans and machines by introducing emotional intelligence, from speech, into conversations with AI. I was building speech models and working on MLOps.

Words people
never use —
could be
only I
know them.
Ishikawa Takuboku 1885–1912

news

Apr 21, 2026	New preprint out! “Where does output diversity collapse in post-training?”, joint work with Xingwei Tan and Nikos Aletras.

selected publications

2026

Where does output diversity collapse in post-training?

Constantinos Karouzos, Xingwei Tan, and Nikolaos Aletras

Apr 2026

Abs HTML

Post-trained language models produce less varied outputs than their base counterparts. This output diversity collapse undermines inference-time scaling methods that rely on varied samples, and risks homogenizing model outputs on creative and value-laden tasks. Prior work attributes collapse to specific post-training methods, without separating the role of training data composition from the method, or the generation format from the model weights. We trace output diversity through three parallel post-training lineages of Olmo 3, Think (chain-of-thought distillation), Instruct (broad multi-source data), and RL-Zero, across 15 tasks and four text diversity metrics. We find that the location of collapse co-varies with data composition: the Think lineage loses most semantic diversity at supervised fine-tuning, and the effect of DPO is larger in Instruct than in Think. Suppressing chain-of-thought reasoning at inference in Think models drops accuracy on hard tasks, yet leaves answer-level diversity unchanged, showing that the collapse is embedded in the model weights by training data, not imposed by the generation format. Decomposing diversity loss on six verifiable tasks into a quality-control component (removal of incorrect outputs) and a residual component (genuine narrowing among correct outputs) reveals that the split is task-dependent, and Think models retain more correct-answer diversity than Instruct despite collapsing more in aggregate. Our results indicate that diversity collapse is determined during training by data composition and cannot be addressed at inference time alone.
An Empirical Study on Preference Tuning Generalization and Diversity Under Domain Shift

Constantinos Karouzos, Xingwei Tan, and Nikolaos Aletras

Jan 2026

Abs HTML

Preference tuning aligns pretrained language models to human judgments of quality, helpfulness, or safety by optimizing over explicit preference signals rather than likelihood alone. Prior work has shown that preference-tuning degrades performance and reduces helpfulness when evaluated outside the training domain. However, the extent to which adaptation strategies mitigate this domain shift remains unexplored. We address this challenge by conducting a comprehensive and systematic study of alignment generalization under domain shift. We compare five popular alignment objectives and various adaptation strategies from source to target, including target-domain supervised fine-tuning and pseudo-labeling, across summarization and question-answering helpfulness tasks. Our findings reveal systematic differences in generalization across alignment objectives under domain shift. We show that adaptation strategies based on pseudo-labeling can substantially reduce domain-shift degradation.