Special Topics: Large Language Models
A research seminar on the papers that introduced and shaped large language models, from the transformer to RLHF, followed by an original student research project on LLMs.
This is a seminar on reading and discussing the papers that introduced and analyzed the important LLMs. In the first half, everyone reads the same papers each week, writes a response, and finds a follow-up paper of their own that pushes the idea toward their research area. In the second half, students propose, carry out, and write up an original research project that studies or applies LLMs. The goal is research initiation: each student leaves with a paper-shaped project in their own field.
Students will:
- understand the transformer architecture and its encoder-only and decoder-only variants;
- find papers that extend the well-known results toward their own research area;
- propose and carry out an original publication about LLMs.
Format. Weekly synchronous discussion (Fridays), led by students, with 5–8 hours a week of independent reading. Generative AI is part of the course: students use LLMs throughout, though paper responses must be their own writing.
Grading. Paper responses 40% · discussion participation and leadership 20% · conference-style paper presentation 5% · research project 35%.
The papers below are the core of my LLM foundations reading list, which is a good place to start if you’re learning how modern language models came to be.
Schedule 13
Week 1
Foundations of LLMs
Fri, Jan 23
transformersattention
- Attention Is All You Need Vaswani et al., 2017 · NeurIPS
Due: paper response and a self-selected follow-up paper
Week 2
Basic models: encoder-only
Fri, Jan 30
BERTpre-training
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding Devlin et al., 2019 · NAACL
Due: paper response and a self-selected follow-up paper
Week 3
Basic models: decoder-only
Fri, Feb 6
GPTgenerative pre-training
- Improving Language Understanding by Generative Pre-Training Radford et al., 2018 · OpenAI
Due: paper response and a self-selected follow-up paper
Week 4
Capabilities: zero-shot learning and reasoning
Fri, Feb 13
GPT-2chain of thoughtprompting
- Language Models are Unsupervised Multitask Learners Radford et al., 2019 · OpenAI
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models Wei et al., 2022 · NeurIPS
Due: paper response and a self-selected follow-up paper
Week 5
Capabilities: retrieval-augmented generation
Fri, Feb 20
RAGretrieval
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Lewis et al., 2020 · NeurIPS
Due: paper response and a self-selected follow-up paper
Week 6
Improvements & training: efficient fine-tuning
Fri, Feb 27
intrinsic dimensionLoRAfine-tuning
- Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning Aghajanyan, Zettlemoyer & Gupta, 2021 · ACL
- LoRA: Low-Rank Adaptation of Large Language Models Hu et al., 2022 · ICLR
Due: paper response and a self-selected follow-up paper
Week 7
Improvements & training: learning from human preferences
Fri, Mar 6
RLHFDPOalignment
- Training Language Models to Follow Instructions with Human Feedback Ouyang et al., 2022 · NeurIPS
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model Rafailov et al., 2023 · NeurIPS
Due: paper response and project proposal
Week 8
Improvements & training: rotary position embeddings
Fri, Mar 13
RoPEpositional encoding
- RoFormer: Enhanced Transformer with Rotary Position Embedding Su et al., 2021 · arXiv
Due: paper response and peer feedback on project proposals
Week 10
Project proposal presentations
Fri, Mar 27
research projects
3 slides from each presenter (week 9 was spring break)
Week 12
New and deep topics: student-chosen papers
Fri, Apr 10
student presentations
Two conference-style presentations of student-chosen papers. Due: paper response, 5 slides per presenter (week 11 was Good Friday)
Week 13
New and deep topics: student-chosen papers
Fri, Apr 17
student presentations
Two conference-style presentations of student-chosen papers. Due: paper response, 5 slides per presenter
Week 14
New and deep topics: student-chosen papers
Fri, Apr 24
student presentations
Two conference-style presentations of student-chosen papers. Due: paper response, 5 slides per presenter
Week 15
Final project presentations
Fri, May 1
research projects
5 slides from each presenter