Prerequisites
Strong mathematical maturity (real analysis, optimization, information theory). Deep learning fundamentals (backprop, CNNs, RNNs). Experience reading ML papers. Recommended: one graduate ML course.
Course D: Graduate Research
Focus: Training, alignment, scaling, interpretability, and frontier topics. Students leave prepared to conduct original research in LLM science. This pathway is the most technically demanding. It opens with a deep dive into attention and decoding because graduate researchers need to reason about architectural modifications at the level of individual operations.
14-Week Syllabus
| Week | Topics | Lab / Assignment |
|---|---|---|
| 1 | Attention and Transformer (deep dive, Ch 03 through 04) | Implement multi-head attention with rotary embeddings |
| 2 | Decoding and Search Algorithms | Implement beam search, MCTS for LLM reasoning |
| 3 | Pre-training and Scaling Laws | Reproduce Chinchilla scaling law predictions |
| 4 | Model Architectures (MoE, SSMs) and Reasoning Models (Ch 07 through 08) | Paper reading: compare MoE routing strategies and reasoning model architectures |
| 5 | Inference Optimization | Implement speculative decoding; analyze test-time compute tradeoffs |
| 6 | Interpretability | Run sparse autoencoder probes on a language model |
| 7 | Synthetic Data and Curriculum Design | Design a synthetic data pipeline for a research task |
| 8 | Fine-Tuning and PEFT (Ch 14 through 15) | Ablation study: rank, target modules, learning rate |
| 9 | Distillation and Model Merging | Distill a large model; merge adapters with TIES/DARE |
| 10 | Alignment (RLHF, DPO, Constitutional AI) | Train a reward model and run DPO |
| 11 | Agents, Tool Use, and Multi-Agent Systems (Ch 22 through 24) | Build an agent with reflection and self-critique |
| 12 | Multimodal Models | Fine-tune a vision-language model |
| 13 | Emerging Architectures and AI and Society (Ch 34 through 35) | Write a research proposal on an open problem |
| 14 | Final project presentations | Novel research contribution (individual or pair) |
Recommended Appendices
- Appendix D: Environment Setup – set up your development environment before Week 1
- Appendix K: HuggingFace: Transformers, Datasets, and Hub – access pretrained models and datasets for labs
- Appendix A: Math Foundations – review the linear algebra and probability behind attention
- Appendix R: Experiment Tracking – log experiments systematically for reproducible research
What Comes Next
Return to the Course Syllabi overview to explore other courses and reading tracks, or proceed to FM.4: How to Use This Book for a quick orientation on conventions and callout types.