Research

We advance open, responsible AI research in Greece and Europe. Our work spans LLMs, ASR, TTS, RAG, and AI safety.

FEATURED RESEARCH · LLM
arXiv preprint

Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

Examines how supervised fine-tuning and reinforcement learning shape Greek reasoning, grammaticality, and instruction following, using behavioral measures beyond answer accuracy.

Read Paper
Greek promptΕξήγησε βήμα προς βήμα.Stage 1SFTbuildsStage 2RLfixesBehavioral measuresReasoningGrammaticalityInstruction followingAnswer accuracyWhat accuracy cannot seeGreek promptΕξήγησε βήμα προς βήμα.Stage 1SFTbuildsStage 2RLfixesBehavioral measuresReasoningGrammaticalityInstruction followingAnswer accuracyWhat accuracy cannot see

Research Areas

Alignment

Making models helpful, honest, and harmless through rigorous alignment and evaluation.

ASR & Speech

Advancing automatic speech recognition and speech understanding for Greek and beyond.

LLM Research

Building frontier language models with a focus on multilingual capability, reasoning, and efficiency.

RAG Systems

Designing robust retrieval systems that ground models in trustworthy, up-to-date knowledge.

AI Safety

Evaluating risks, improving robustness, and ensuring responsible deployment at scale.

Latest Research

4 papers · arXiv preprints