Welcome to the course!

This is a seminar course on modern language modeling. We study language models from four perspectives: how they are trained (data, optimization, and training stability), how they generate (autoregressive vs. diffusion decoding), what they compute (mechanistic and actionable interpretability; the expressivity of Transformers vs. RNNs), and how we measure progress (the methodology---and taste---of evaluation). Meetings are a 3-hour weekly discussion seminar centered on student-led paper presentations, plus the presentation and discussions from CompLING lab members.

People

• Instructor: Freda Shi
• Discussion forum: Piazza (questions should go through the forum first)

Time and Location

• Weekly seminar meeting: Friday 9:00AM-11:50AM, DC 2568

Textbook

There is no official textbook. Readings consist of research papers, listed per week in the schedule below. Everyone is expected to familiarize themselves with the required papers before class; the optional list is mainly for presenters and for those who want to go deeper.

Tentative Course Schedule and Related Material

The seminar meets once a week for a 3-hour session: brief context setting, student-led paper presentations (each student presents twice over the term, 20-30 minutes each including Q&A), and an open discussion and synthesis block. First week of classes: September 9–11 (UW classes begin Wednesday, September 9). No meeting during Reading Week (October 10–18). Presenter sign-up opens in Week 1.

Presentation sign-up policy.

• Everyone ranks the twelve weekly topics (Lectures 1-12) by preference and submits the ranking through this Google Form, due by the end of Friday, September 11.
• Your first topic is assigned greedily in a randomized order: the class roster, sorted by (last name, first name), is randomly shuffled with seed 42 in this Python script. The second topic is assigned greedily in the reverse of that order.
• Authors cannot sign up for a session covering their own paper.
• Members of each weekly group coordinate among themselves on who presents which part, but every presenter must be familiar with the details of all required readings for that week.
• During the discussion block, the instructor will randomly call on presenters to answer questions about the readings, which counts for 5 of the 20 percentage points of that presentation.

Week	Topic and Readings	Presenter(s)
1	Sep 11 Foundations: the Transformer and autoregressive language modeling Self introduction; course logistics. Focus: self-attention and positional structure; what next-token prediction means; anatomy of a pretraining run. Required readings: Vaswani et al., NeurIPS 2017: Attention Is All You Need Radford et al., 2019: Language Models are Unsupervised Multitask Learners (the GPT-2 report) Optional readings: Brown et al., NeurIPS 2020: Language Models are Few-Shot Learners (GPT-3) Alammar (blog walk-through): The Illustrated Transformer	Freda Shi (instructor lecture)
2	Sep 18 Pretraining data science I: scaling laws (how much data, and of what kind) Focus: compute-optimal allocation between parameters and tokens; what changes when data, not compute, is the bottleneck. Required readings: Kaplan et al., 2020: Scaling Laws for Neural Language Models Hoffmann et al., NeurIPS 2022: Training Compute-Optimal Large Language Models (Chinchilla) Muennighoff et al., NeurIPS 2023: Scaling Data-Constrained Language Models Chudnovsky et al., 2026: Internal Data Repetition Destroys Language Models Optional readings: Caballero et al., ICLR 2023: Broken Neural Scaling Laws Villalobos et al., 2022: Will We Run Out of Data? An Analysis of the Limits of Scaling Datasets in Machine Learning	Hrutil Patel, Krish Patel, Clément Salomon
3	Sep 25 Pretraining data science II: curation (filtering, deduplication, and data mixing) Focus: data quality vs. quantity; controlled data ablations as an experimental science. Required readings: Penedo et al., NeurIPS 2023: The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only Li et al., NeurIPS 2024: DataComp-LM: In Search of the Next Generation of Training Sets for Language Models Optional readings: Xie et al., NeurIPS 2023: DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining Abbas et al., 2023: SemDeDup: Data-Efficient Learning at Web-Scale through Semantic Deduplication Gunasekar et al., 2023: Textbooks Are All You Need (phi-1) Soldaini et al., 2024: Dolma: An Open Corpus of Three Trillion Tokens for Language Model Pretraining Research	Idil Kara, Soheil Soltani
4	Oct 2 Optimization I: adaptive methods and gradient instability Focus: why gradients explode, why large LMs need tame updates; clipping as a form of adaptivity. Required readings: Kingma & Ba, ICLR 2015: Adam: A Method for Stochastic Optimization Pascanu, Mikolov & Bengio, ICML 2013: On the Difficulty of Training Recurrent Neural Networks Zhang et al., ICLR 2020: Why Gradient Clipping Accelerates Training: A Theoretical Justification for Adaptivity Optional readings: Loshchilov & Hutter, ICLR 2019: Decoupled Weight Decay Regularization (AdamW) Molybog et al., 2023: A Theory on Adam Instability in Large-Scale Machine Learning	Hrutil Patel, Krish Patel, Hella Pimenova
5	Oct 9 Optimization II: stability at LLM scale and beyond-Adam optimizers Focus: loss spikes and stability engineering in practice; Muon and preconditioned updates as geometry-aware stabilization. Required readings: Liu et al. (Moonshot AI), 2025: Muon is Scalable for LLM Training; see also Muon (Jordan et al., 2024) Groeneveld et al. (AI2), 2025: OLMo 2 Furious Gupta, Koren & Singer, ICML 2018: Shampoo: Preconditioned Stochastic Tensor Optimization Optional readings: Chowdhery et al., 2022: PaLM: Scaling Language Modeling with Pathways Yang et al., ICLR 2022: Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer (μTransfer) Vyas et al., 2024: SOAP: Improving and Stabilizing Shampoo using Adam Bernstein & Newhouse, 2024: Old Optimizer, New Norm: An Anthology	Calder Johnson, Zuoqiu (Robert) Liu, Hella Pimenova
Reading Week (Oct 10–18): No class
6	Oct 23 Diffusion language models Focus: iterative denoising vs. left-to-right decoding; discrete diffusion objectives; parallel generation and controllability. Required readings: Austin et al., NeurIPS 2021: Structured Denoising Diffusion Models in Discrete State-Spaces (D3PM) Lou, Meng & Ermon, ICML 2024: Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution (SEDD) Nie et al., 2025: Large Language Diffusion with mAsKing (LLaDA) Optional readings: Li et al., NeurIPS 2022: Diffusion-LM Improves Controllable Text Generation Inception Labs, 2025: Mercury: Ultra-Fast Language Models Based on Diffusion Gemini Diffusion (Google DeepMind, 2025; product/tech preview)	Xiaoxi Luo, Salman
7	Oct 30 From autoregressive to diffusion: unifying views and hybrids Focus: any-order generation as a generalization of AR decoding; block-wise interpolation between AR and diffusion. Required readings: Sahoo et al., NeurIPS 2024: Simple and Effective Masked Diffusion Language Models (MDLM) Arriola et al., ICLR 2025: Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models Zhou et al., 2024: Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model Optional readings: Han et al., ACL 2023: SSD-LM: Semi-Autoregressive Simplex-Based Diffusion Language Model for Text Generation and Modular Control Chen et al., NeurIPS 2024: Diffusion Forcing: Next-Token Prediction Meets Full-Sequence Diffusion	Henry Guo, Salman
8	Nov 6 Mechanistic interpretability I: circuits and superposition Focus: reverse-engineering small Transformers; induction heads; why features are hard to read and how sparse coding helps. Required readings: Elhage et al., 2021: A Mathematical Framework for Transformer Circuits Olsson et al., 2022: In-context Learning and Induction Heads Elhage et al., 2022: Toy Models of Superposition Optional readings: Bricken et al., 2023: Towards Monosemanticity: Decomposing Language Models With Dictionary Learning Cunningham et al., ICLR 2024: Sparse Autoencoders Find Highly Interpretable Features in Language Models	Calder Johnson, Asim Waheed
9	Nov 13 Interpretability II: actionable interpretability, and vision-language models Focus: from understanding to control (feature steering, editing); what carries over to VLMs. Required readings: Templeton et al., 2024: Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet Turner et al., 2023: Activation Addition: Steering Language Models Without Optimization Goh et al., Distill 2021: Multimodal Neurons in Artificial Neural Networks Wu et al., ICML 2026: The Mechanistic Emergence of Symbol Grounding in Language Models Optional readings: Wang et al., ICLR 2023: Interpretability in the Wild: A Circuit for Indirect Object Identification in GPT-2 Small Meng et al., NeurIPS 2022: Locating and Editing Factual Associations in GPT (ROME) Gandelsman et al., ICLR 2024: Interpreting CLIP’s Image Representation via Text-Based Decomposition	Idil Kara, Soheil Soltani, Asim Waheed
10	Nov 20 Expressivity: Transformers vs. RNNs, and the power of next-token prediction Focus: log-precision attention sits in TC0; chain-of-thought changes the complexity class; what the next-token distribution can represent; do matrix-valued recurrent states escape TC0? Required readings: Hahn, TACL 2020: Theoretical Limitations of Self-Attention in Neural Sequence Models Merrill & Sabharwal, ACL 2022: The Parallelism Tradeoff: Limitations of Log-Precision Transformers to Constant-Depth Threshold Circuits Merrill & Sabharwal, ICLR 2024: The Expressive Power of Transformers with Chain of Thought Wang, Zhu & Shi, 2025: Distribution Prompting: Understanding the Expressivity of Language Models Through the Next-Token Distributions They Can Produce Mishra et al., 2026: M²RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling Optional readings: Merrill, Petty & Sabharwal, ICML 2024: The Illusion of State in State-Space Models Pérez, Barceló & Marinković, JMLR 2021: Attention is Turing-Complete	Kath Choi, Zuoqiu (Robert) Liu, Xiaoxi Luo, Clément Salomon
11	Nov 27 LM evaluation I: benchmarks, leaderboards, and how they break Focus: what HELM set out to fix; saturated and contaminated benchmarks; human preference leaderboards and their failure modes. Required readings: Liang et al., TACL 2023: Holistic Evaluation of Language Models (HELM) Gema et al., 2024: Are We Done with MMLU? (MMLU-Redux) Chiang et al., ICML 2024: Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference Optional readings: Dehghani et al., 2021: The Benchmark Lottery Reuel et al., 2024: BetterBench: Assessing AI Benchmarks, Uncovering Issues, and Establishing Best Practices Singh et al., 2025: The Leaderboard Illusion	Max Homm, Bihui Jin
12	Dec 4 LM evaluation II: frontier benchmarks and the case for “taste” Focus: what would count as “solved”; why Humanity’s Last Exam is not the end of evaluation; why ARC(-AGI) still is not enough. Required readings: Clark et al., 2018: Think You Have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge Chollet, 2019: On the Measure of Intelligence; and Chollet et al., 2025: ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems Phan et al., 2025: Humanity’s Last Exam (HLE) Mirzadeh et al., 2024: GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models Optional readings: Rein et al., 2023: GPQA: A Graduate-Level Google-Proof Q&A Benchmark Schaeffer et al., NeurIPS 2023: Are Emergent Abilities of Large Language Models a Mirage? Gao et al., ICML 2023: Scaling Laws for Reward Model Overoptimization	Kath Choi, Henry Guo, Max Homm, Bihui Jin

Grade Breakdown

• Paper presentations: 2 x 20%
  • 2 x 5% of the presentation grade comes from the random cold-call Q&A during discussion.
  • Signing up to present in Lecture 2 or Lecture 3 receives an overall bonus of 5% and 2.5%, respectively.
• Class participation and discussion: 10%
  • Graded by how actively you participate in person; attending every session guarantees a baseline of 7%.
• Paper critiques: 25% (five critiques, 5% each)
  • Pick five papers from weeks in which you are not presenting, and post one critique per paper directly on the Piazza forum (no separate submission is needed).
  • A critique counts only if it is posted by 23:59 on the day before the corresponding topic’s session.
  • Submitting earns 1%; the rest rewards constructiveness (longer is not better).
  • AI use is discouraged. A critique found to contain factual errors or meaningless questions is capped at 3% (out of the 5% for that critique).
• Course project: 25%
  • Any research project on NLP, preferably related to the course content. If you are unsure about the topic, you are welcome to discuss it with the instructor after class.
  • Deliverables: a 4-page report in ACL format plus your project code, submitted on LEARN, due Tuesday, December 8.

Course Policies

We follow the University of Waterloo Academic Integrity policy.