Trustworthy Large Language Models
Semester I, 2026-27
Updates
- New Lecture is up: 22. Hallucination in LLMs - I: Why LLMs Hallucinate? [ slides ]
- New Lecture is up: 21. Uncertainty in LLMs - IV: Accountable LLM Systems [ slides ]
- New Lecture is up: 20. Uncertainty in LLMs - III: Safe Non-Answering [ slides ]
- New Lecture is up: 19. Uncertainty in LLMs - II: Semantic and Claim-Level Risk [ slides ]
- New Lecture is up: 18. Uncertainty in LLMs - I: Trustworthy Confidence [ slides ]
- New Lecture is up: 17. Watermarking - II [ slides ]
- New Lecture is up: 16. Watermarking - I [ slides ]
Course Description
Large Language Models (LLMs) are rapidly becoming the foundation of modern AI systems, powering applications ranging from conversational assistants and code generation to scientific discovery and autonomous agents. While recent advances have dramatically improved their capabilities, deploying these models responsibly requires addressing fundamental challenges related to privacy, security, robustness, fairness, reliability, and interpretability. This advanced course provides a comprehensive introduction to the principles, techniques, and emerging research behind trustworthy and responsible LLMs. Beginning with the foundations of transformer-based language models, instruction tuning, alignment, and decoding strategies, students will develop the background necessary to understand how modern LLMs are trained and deployed.
This course then explores the key challenges that arise in real-world LLM systems, including privacy leakage, adversarial attacks, jailbreaking, hallucinations, uncertainty estimation, algorithmic bias, watermarking, and mechanistic interpretability. Alongside state-of-the-art defense mechanisms such as differential privacy, robust training, retrieval-augmented generation, verification, calibration, and machine unlearning, students will examine recent research papers, benchmarks, and practical techniques for building trustworthy AI systems. By the end of the course, students will have a deep understanding of both the capabilities and limitations of contemporary language models, equipping them to critically evaluate, design, and develop the next generation of reliable, secure, and accountable AI systems.
Students should have a background in Machine Learning and be proficient in Python programming. At least some basic knowledge of Deep Learning and NLP is preferred. Through assignments and a course project, students will acquire the skills necessary to design, implement, and understand LLMs using the PyTorch framework.
Previous Offerings
Logistics
- Timings: Slot AA (Monday, Thursday: 02:00 -- 03:30 PM)
- Classroom: Bharti-301
- News and announcements: All the news and announcements will be posted on Piazza.
- Piazza link: Sign up for this
class on Piazza. Check your mail for the access code.
Students are encouraged to ask questions and participate in discussions! - Assignment submission: All assignments should be submitted on Moodle.
Assessment Plan (Tentative)
- Minor Exam - 15%
- Major Exam - 25%
- Quizzes (3) - 15% (5% Each)
- Assignments (2) - 20% (10% Each)
- Course Project (Group-wise) - 25%