• Foundations of Large Language Models – Transformers and self-attention, Pre-training and scaling, Instruction fine-tuning, Alignment and preference optimization, Decoding strategies

  • Privacy in Large Language Models – Privacy fundamentals, Text sanitization, Memorization and data extraction, Membership inference attacks, Prompt and in-context learning leakage, Prompt injection attacks, Differential privacy

  • Security and Robustness – Adversarial attacks in continuous and discrete spaces, Gradient-based discrete optimization, Jailbreaking techniques, Transferability of adversarial attacks, Safety training failures, Training data poisoning, Alignment faking, Evaluation awareness, Chain-of-thought faithfulness, Indirect prompt injection, Memory poisoning in agentic systems, Monitoring and verification, Automated red teaming, Policy languages and compositional verification

  • Bias and Fairness – Foundations of fairness, Social and algorithmic bias, Fairness metrics and benchmarks, Bias in pre-training, tokenization, and RLHF, Debiasing techniques, Limitations of current approaches, Policy and governance considerations

  • Watermarking – Motivation and applications, Properties of effective watermarks, Formal problem formulation, Red–Green watermarking, Training-time and decoding-time watermarking, Threat models and attacks

  • Hallucination and Factuality – Causes of hallucinations, Hallucination detection benchmarks, Evaluation metrics, Decoding-based mitigation, Post-hoc verification, Self-correction, Retrieval-Augmented Generation (RAG)

  • Uncertainty and Trustworthy Generation – Confidence estimation and calibration, Self-knowledge in language models, Semantic uncertainty, Unsupported claim detection, Claim-level risk estimation, Abstention and deferral, Risk-controlled generation, Evidence grounding, Auditability and oversight, Grounded and governed LLM systems, LLM judges, RAG, multimodal systems, and AI agents

  • Interpretability – Attention and feature attribution, Mechanistic interpretability, Probing methods, Logit Lens and Patchscopes, Activation patching, Neuron-level analysis, Circuits and causal abstraction, Polysemanticity, Linear representation hypothesis, Dictionary learning, Transcoders and cross-coders, Model diffing, Activation steering, Interpretability for alignment and safety