A new PhD thesis proposes a three-part system to make large language models hallucinate less, block harmful output as it's generated, and adapt to different languages and cultures instead of applying one rulebook everywhere.
The thesis was published as a preprint on September 23, 2026, and frames itself as a blueprint for what it calls responsible intelligence. It tackles three problems separately: reducing hallucinations in specialized domains through active learning and graph-based knowledge, and blocking harmful text mid-generation through a new decoding-time alignment mechanism rather than filtering output after the fact. The third piece applies language-specific steering to multilingual models, aiming to respect different linguistic and social norms instead of enforcing one global standard.
Most AI safety work today happens at training time, through reinforcement learning from human feedback, or after the fact, through output filters layered on top of a finished model. Catching problems as the model is still writing, and tuning that catch separately for each domain and language, is a more granular approach than either of those on paper. Whether it holds up outside a thesis defense is a different question.
Calling a research proposal a blueprint for next-generation AI is the kind of line that reads better in an abstract than it will in a benchmark.