Schedule

Course Overview and Introduction to LLMs

One-off: this week meets twice, Monday and Thursday, 14:30–16:00, in Classroom 3223. From week 2 the course meets Mondays only, 14:30–17:30, in Classroom 4225.

1. Mon 31 Aug

   Introduction & Course Overview

   Slides ↓

2. Thu 3 Sep

   N-gram Language Models

   Slides ↓

   Further reading N-gram Language Models Daniel Jurafsky & James H. Martin Speech and Language Processing, 3rd ed. draft (Aug 2026), ch. 3 PDF ↓ Free textbook chapter behind the N-gram lecture.

Tokenization and Vector Embeddings; Neural Language Models

1. Mon 7 Sep

   Slides: Tokenization and Embeddings ↓ Slides: Neural Language Models ↓

   Required A Neural Probabilistic Language Model Bengio, Ducharme, Vincent & Jauvin JMLR 2003

   Further reading Words and Tokens Jurafsky & Martin Speech and Language Processing, 3rd ed. draft (Aug 2026), ch. 2 PDF ↓ Background for the tokenization half of the lecture.

   Further reading Embeddings Jurafsky & Martin Speech and Language Processing, 3rd ed. draft (Aug 2026), ch. 5 PDF ↓ Background for the vector-semantics half of the lecture.

Attention and Transformers

1. Mon 14 Sep

   Attention

   Transformers

   Slides: Attention ↓ Slides: Transformers ↓

   Required Attention Is All You Need Vaswani et al. NeurIPS 2017

   Further reading Neural Networks Jurafsky & Martin Speech and Language Processing, 3rd ed. draft (Aug 2026), ch. 6 PDF ↓ A refresher, if you need one.

   Further reading Transformers and Pretraining Jurafsky & Martin Speech and Language Processing, 3rd ed. draft (Aug 2026), ch. 7 PDF ↓ Textbook companion to Attention Is All You Need.

Post-training, Alignment, and Tool Use

1. Mon 21 Sep

   Slides ↓

   Required ReAct: Synergizing Reasoning and Acting in Language Models Yao et al. ICLR 2023

   Required s1: Simple test-time scaling Muennighoff et al. EMNLP 2025

   Further reading Post-training Jurafsky & Martin Speech and Language Processing, 3rd ed. draft (Aug 2026), ch. 8 PDF ↓ Instruction tuning, parameter-efficient fine-tuning and preference alignment.

   Further reading Tree of Thoughts: Deliberate Problem Solving with Large Language Models Yao et al. NeurIPS 2023 Search over a tree of intermediate thoughts instead of one chain.

   Further reading Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? Yue et al. NeurIPS 2025 At high pass@k, RLVR models do not beat their own base model.

Hallucinations: Why They Happen, and How to Catch Them

1. Mon 28 Sep

   Slides ↓

   Required Detecting hallucinations in large language models using semantic entropy Farquhar, Kossen, Kuhn & Gal Nature 2024

   Then the paper for your track:

   Track A Why Language Models Hallucinate Kalai et al. arXiv 2025

   Track B Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations Sun et al. NeurIPS 2025

   Further reading Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools Magesh et al. Journal of Empirical Legal Studies, 2025 Legal RAG products advertised as hallucination-free hallucinate 17–33% of the time.

   Further reading AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions Kirichenko et al. NeurIPS 2025 Reasoning fine-tuning makes models answer confidently when they should abstain.

Evaluating LLMs: Benchmarks, Judges and Leaderboards

1. Mon 5 Oct

   Slides ↓

   Required Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs Betley et al. ICML 2025

   Then the paper for your track:

   Track A Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference Chiang et al. ICML 2024

   Track B HALoGEN: Fantastic LLM Hallucinations and Where to Find Them Ravichander et al. ACL 2025

   Further reading Survey on Evaluation of LLM-based Agents Yehudai et al. arXiv 2025 Survey of agentic, trajectory and environment-based evaluation.

   Further reading Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena Zheng et al. NeurIPS 2023 Position, verbosity and self-enhancement bias in LLM judges.

   Further reading The Leaderboard Illusion Singh et al. NeurIPS 2025 Undisclosed private testing and selective disclosure skew Arena rankings.

RAG at Scale and Multi-Agent Frameworks

1. Mon 12 Oct

   Slides ↓

   Required Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Lewis et al. NeurIPS 2020

   Required AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation Wu et al. COLM 2024

   Further reading Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection Asai et al. ICLR 2024 Reflection tokens let the model decide when to retrieve and critique what comes back.

   Further reading The Faiss library Douze et al. IEEE Transactions on Big Data, 2026 The indexing and quantization machinery under retrieval.

   Further reading Retrieval meets Long Context Large Language Models Xu et al. ICLR 2024 4K context plus retrieval matches a fine-tuned 16K model.

Mid-semester Break

1. Mon 19 Oct

   Mid-semester break

   No class

Agent Infrastructure, Memory, and Observability

1. Mon 26 Oct

   Required Why Do Multi-Agent LLM Systems Fail? Cemri et al. NeurIPS 2025

   Then the paper for your track:

   Track A MemGPT: Towards LLMs as Operating Systems Packer et al. arXiv 2023

   Track B Recursive Language Models Zhang, Kraska & Khattab arXiv 2025

   Further reading Lost in the Middle: How Language Models Use Long Contexts Liu et al. TACL 2024 Models use the start and end of a long context and lose the middle.

   Further reading Model Context Protocol: Architecture Anthropic MCP documentation Protocol documentation, not a paper. No summary due.

   Further reading What is A2A? A2A Project A2A protocol documentation Protocol documentation, not a paper. No summary due.

LLM System Design and Serving at Scale

1. Mon 2 Nov

   Required Orca: A Distributed Serving System for Transformer-Based Generative Models Yu et al. OSDI 2022

   Required Efficient Memory Management for Large Language Model Serving with PagedAttention Kwon et al. SOSP 2023

   Further reading Mooncake: Trading More Storage for Less Computation — A KVCache-centric Architecture for Serving LLM Chatbot Qin et al. FAST 2025 Disaggregated serving built around the KV cache; in production behind Kimi.

   Further reading Taming the Titans: A Survey of Efficient LLM Inference Serving Zhen et al. arXiv 2025 Survey of speculative decoding, quantization and inference scheduling.

   Further reading RouteLLM: Learning to Route LLMs with Preference Data Ong et al. ICLR 2025 Routing between a strong and a weak model: >2× cheaper at matched quality.

LLMs for Scientific Discovery

1. Mon 9 Nov

   Which two papers you read depends on your track:

   Track A Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers Si, Yang & Hashimoto ICLR 2025

   Track A The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas Si, Yang & Hashimoto arXiv 2025

   Track B Barbarians at the Gate: How AI is Upending Systems Research Cheng et al. arXiv 2025

   Track B The Last Human-Written Paper: Agent-Native Research Artifacts Liu et al. arXiv 2026

   Further reading Predicting Empirical AI Research Outcomes with Language Models Wen et al. NeurIPS 2025 Forecasting which experiments will pan out before running them.

   Further reading PaperBench: Evaluating AI's Ability to Replicate AI Research Starace et al. ICML 2025 Can agents reproduce 20 ICML papers from scratch? Rubrics by the original authors.

LLMs in Healthcare and Education

1. Mon 16 Nov

   Which two papers you read depends on your track:

   Track A Towards conversational diagnostic artificial intelligence Tu et al. Nature 2025

   Track A An evaluation framework for clinical use of large language models in patient interaction tasks Johri et al. Nature Medicine 2025

   Track B AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting Kestin et al. Scientific Reports 2025

   Track B MathTutorBench: A Benchmark for Measuring Open-ended Pedagogical Capabilities of LLM Tutors Macina et al. EMNLP 2025

   Further reading Towards conversational artificial intelligence for disease management Liévin, Palepu, Weng et al. Nature 2026 AMIE across multiple visits, grounded in clinical guidelines; non-inferior to 21 physicians.

   Further reading Large language models for frontline healthcare support in low-resource settings Rutunda et al. Nature Health 2026 101 community health workers in Rwanda: the models beat the local clinicians.

   Further reading Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise Wang, Ribeiro, Robinson, Loeb, Demszky arXiv 2024 Give the model to the tutor, not the student: +9 p.p. for the weakest tutors.

From Language Models to World Models

1. Mon 23 Nov

   Required Language Models Represent Space and Time Gurnee & Tegmark ICLR 2024

   Required Evaluating the World Model Implicit in a Generative Model Vafa et al. NeurIPS 2024

   Further reading The Platonic Representation Hypothesis Huh et al. ICML 2024 Vision and language converge on one model of reality.

   Further reading Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task Li et al. ICLR 2023 A GPT trained only on Othello moves builds a board state you can intervene on.

   Further reading V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning Assran et al. arXiv 2025 Learn the world from video, not text. Not a language model.

AI Agents in the Physical World

1. Mon 30 Nov

   Required OpenVLA: An Open-Source Vision-Language-Action Model Kim et al. CoRL 2024

   Required π₀: A Vision-Language-Action Flow Model for General Robot Control Black et al. RSS 2025

   Further reading Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer Gemini Robotics Team, Google DeepMind arXiv 2025 Generality from a frontier VLM's world knowledge rather than robot data.

   Further reading π₀.₅: A Vision-Language-Action Model with Open-World Generalization Physical Intelligence CoRL 2025 Does the recipe transfer to homes it has never seen?

   Further reading Open X-Embodiment: Robotic Learning Datasets and RT-X Models Open X-Embodiment Collaboration arXiv 2023 1M+ trajectories, 22 robots, 527 skills: the cross-embodiment dataset.

Project Presentations

1. Mon 7 Dec

   Project presentations