Why Do Prompts Fail: The Hidden Mechanics of LLMs

Large Language Models (LLMs) have transformed how we interact with AI, enabling natural language to power tasks from content generation to complex problem solving. However, their behavior can often be unpredictable, where even small changes in prompts lead to errors, inconsistencies, or hallucinations.

In this session, we’ll explore the hidden mechanics behind prompt failures, uncovering how LLMs process and generate language. You’ll gain insights into concepts like probabilistic token prediction, attention mechanisms, and in-context learning, and understand why certain prompts succeed while others fail.

The goal is to move beyond trial-and-error prompting and develop a more systematic, reliable approach to designing prompts that produce accurate and consistent results.

This is an insight-driven session designed to help practitioners build better intuition and control over LLM behavior.

Key Takeaways:

• Why Prompts Fail – understanding inconsistencies and hallucinations in LLM outputs
• Inner Workings of LLMs – token prediction, attention, and in-context learning explained
• Prompt Design Principles – moving from trial-and-error to structured prompting
• Improving Reliability – techniques to generate consistent and accurate outputs
• System Thinking – building intuition for controlling LLM behavior

Dr. Avinash Kumar Singh is a Senior Researcher at Domyn with over 14 years of experience in artificial intelligence, spanning research, engineering, and leadership. His work focuses on large language models, computer vision, and human–AI interaction, with an emphasis on building scalable, real-world AI systems.

He holds a Ph.D. in Computer Vision and has conducted postdoctoral research in human-robot interaction at Umeå University and Montpellier University. He has contributed to leading conferences and journals, and actively mentors professionals in AI. His research interests include Responsible and Sovereign AI, with a focus on improving the reliability and behavior of large language models in real-world deployments.