Erfan Miahi

Introduces CCPS, a novel method for more accurately estimating Large Language Model (LLM) confidence by analyzing internal representational stability. It applies targeted adversarial perturbations to final hidden states, extracts features, and uses a lightweight classifier to predict answer correctness. The paper demonstrates significant improvements in calibration metrics and accuracy over existing methods.

A research paper (pre-print on arXiv) exploring neural network representations within reinforcement learning.

Master's thesis investigating the connection between properties and generalization performance of representations learned by deep reinforcement learning algorithms. It empirically examines representations good at generalization and proposes novel hypotheses regarding the impact of auxiliary tasks in end-to-end training.

Published in Expert Systems and Applications, this paper details the use of Genetic Neural Architecture Search for the automatic assessment of human sperm images.

Published in IEEE Transactions on Cybernetics, this work addresses scalable transfer evolutionary optimization for handling large task instances.

Published in Computers in Biology and Medicine, this paper explores the impact of deep transfer and multi-task learning on detecting sperm abnormalities.