Noor Islam S. Mohammad

Achieving a macro-F1 score of 0.940 reveals that task-aware feature gating can dramatically enhance multi-task NLP performance by addressing class imbalance and inductive bias.

BPF can detect 89.6% of semantic drift in formalizations while reducing erroneous outputs by 47%, transforming the reliability of autoformalization in mathematics.

Chain-of-thought reasoning amplifies anti-Muslim bias in LLMs, revealing critical gaps in existing mitigation strategies across realistic deployment scenarios.

Federated intrusion detection can achieve centralized accuracy with 97% less communication overhead by binarizing gradients and encrypting them homomorphically.