CosmicAI Research, “Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning” selected for a paper award at STAI-X!

Artificial-intelligence systems are often most confident when they have the least reason to be. This tendency can make them unreliable and potentially harmful. 

“Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning,” by Pietro Carlotti, Nevena Gligić and Arya Farahi, tackles that awkward habit. 

The proposed method helps a model judge whether the data in front of it looks familiar or unusual, then adjusts its confidence accordingly. In particular, it encourages an AI system to admit when it is out of its depth rather than make a confident claim without having encountered enough similar data to support its prediction. 

That matters in fields such as medicine, science and automated decision-making, where misplaced certainty can be more dangerous than an honest “I don’t know”. 

The paper was selected for a paper award at STAI-X, reflecting a growing priority in AI research: building systems whose confidence deserves as much scrutiny as their answers.

Pietro Carlotti presenting “Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning”