Notable publication: Why Attention Fails – Faults in Attention-Based Neural Networks

In the paper “Why Attention Fails: A Taxonomy of Faults in Attention-Based Neural Networks,” Sigma Jahan, Saurabh Singh Rajput, Tushar Sharma, and Mohammad Masudur Rahman present the first large-scale empirical study of faults in attention-based neural networks (ABNNs). The study systematically analyses 555 real-world faults drawn from 96 projects across ten frameworks, including GitHub, Hugging Face, and Stack Overflow, to understand how and why attention mechanisms break in practice.

From this analysis, the authors derive a novel taxonomy of seven attention-specific fault categories that are not captured by existing defect studies. Their findings have direct implications for testing, debugging, and hardening attention-based models used in safety- and mission-critical applications. The work was published at the **CORE A*–ranked International Conference on Software Engineering (ICSE)**.

Read the paper: https://arxiv.org/abs/2508.04925