In one line. Mondays are lectures. Discussion sessions use fixed groups of four or five, with an individual preparation note due Tuesday at 18:00. Four hours a week, two of them contact time. No exam.
Contents
- 1. Goal
- 2. Prerequisites
- 3. How the course runs
- 4. Schedule
- 5. The individual paper-search project
- 6. Assessment
- 7. Generative AI policy
- 8. Reading list
- 9. Logistics
1. Goal
By December, students will be able to:
- Place a retrieval method on the spectrum from sparse to dense, and say when BM25 still beats it.
- Say where an AI system’s knowledge sits when it runs, and what that choice costs.
- Compare retrieval methods under a fixed protocol, and explain what the evaluation measures and what it hides.
- Find where a fluent, correct-sounding account of a paper is subtly wrong.
- Report from their own records where language models fail on research reading.
- Diagnose why a change to a retrieval system did or did not help.
- Defend a technical position under challenge, and change it when given a reason.
Objectives 4, 5 and 7 matter most. Few students will ever build a search engine. All of them will read papers whose claims outrun their evidence, work with models that are fluent and wrong, and have to defend a position in a room.
The two threads that run through the term
From sparse to dense. TF-IDF, BM25, learned sparse retrieval, bi-encoders, late interaction. A spectrum, not a succession.
Where knowledge sits when the system runs. One figure, redrawn every few weeks.
| Regime | Mechanism | Example |
|---|---|---|
| Parametric | All knowledge in weights, fixed compute | A dense LLM |
| Conditionally parametric, by compute | A router selects which parameters to run | Mixture of experts |
| Conditionally parametric, by lookup | A hash selects which parameters to read | Product-key memory, memory layers, Engram |
| Written at inference | The memory is learned while the model runs | Titans, cartridges, sparse memory finetuning |
| Semi-parametric | An index wired into the forward pass | kNN-LM, RETRO, memory networks |
| Non-parametric | A corpus placed into the prompt | Retrieval-augmented generation |
| Open world | A policy decides what to fetch | Search agents, deep research |
The course opens with attention itself as retrieval over the context, so the whole spectrum reads as one idea at different scales.
The middle three rows are where the figure changed in 2026. Routing computation and routing a lookup used to be one row, until a result that year found them complementary rather than alternative, with an optimum between them. The row above them did not exist at all.
2. Prerequisites
- Familiarity with Python and PyTorch coding.
- Working knowledge of transformers and standard deep learning.
- Basic maths: linear algebra, probability, statistics.
- No prior background in information retrieval is required, and none is expected. The classical material in September will be the new material for most of this audience.
3. How the course runs
The first four weeks are lectures only. Discussion sessions begin on 30 September, once groups are formed and there is enough retrieval background to argue about a paper rather than decode it.
Monday is a lecture on the method. From session 4, each lecture has 35 minutes of core material, 10 minutes for carry-over or guided practice, 10 minutes of questions and 5 minutes for course business.
Discussion sessions use fixed groups of four or five. The next discussion is BEIR on 30 September. Check the schedule: some Wednesdays are lectures.
Preparation is due Tuesday 18:00. Read the anchor and the relevant sections of one supporting paper you choose. Your assigned role guides your preparation, not the opinion you must defend. The preparation form contains your role, supporting paper, one-sentence argument, 2-3 paragraphs of reasoning and evidence, one paragraph on a remaining question for peers, and an AI declaration.
A new case is revealed at 09:00-09:05. Groups discuss from 09:05-09:30. One or two students per group answer short graded follow-up questions within 09:30-09:40. Each student submits a closing paragraph at 09:40-09:45, before lecturer synthesis. The final three minutes are for a private questionnaire about English and AI use in the group. Groups manage their own discussion time. There is no written group report.
Use English throughout. AI is optional for preparation with disclosure, and is prohibited during the discussion session, including the closing paragraph.
Amended 16 September 2026. This replaces the half-page, four-field review, mandatory model check, timed speaking rounds and ranked group conclusion. The preparation and discussion policy gives the complete form, hour, submission rules and privacy terms. Open BEIR preparation for the 29 September deadline and your assigned role.
4. Schedule
All sessions run 09:00-10:00 KST, including the three project sessions.
Pacing update, 7 September. BM25 finishes in session 4 (9 September); evaluation moves to session 5 (14 September). The web-search survey and three-engine comparison become optional reference and practice. Later lectures focus on the core topics below, with extensions kept in the reading list. Optional material adds no submission or required homework. Discussion dates, project workshops, review deadlines and the 9 December write-up deadline stay as published.
The topic plan will continue to follow the pace of the class. Carry-over uses the reserved time first; further changes will reduce scope before adding workload.
| # | Date | Content | Due |
|---|---|---|---|
| 1 | Mon 31 Aug | Why retrieval now. Eight arguments, the knowledge-location spectrum, course mechanics Deck - Recording | - |
| 2 | Wed 02 Sep | The inverted index. Why grep does not scale, dictionary and postings, merging, phrases, wildcards and the k-gram index Deck - Recording | - |
| 3 | Mon 07 Sep | Index construction and ranking foundations. Index construction carried from the inverted-index lecture, TF-IDF vectors, relevance odds and the binary independence score through probabilistic IDF. Deck - Recording | - |
| 4 | Wed 09 Sep | BM25: finish probabilistic ranking. Probability ranking principle, term-frequency saturation, length normalisation and BM25 parameter experiments. Deck - Recording | - |
| 5 | Mon 14 Sep | Evaluation and the changing goals of retrieval benchmarks. Precision and recall, nDCG by hand, pooling and judgement bias. The evaluation foundations for BEIR and the paper-search project. Deck - Recording | - |
| 6 | Wed 16 Sep | Coarse retrieval and reranking. Candidate generation, ranking, fusion, reranking and practical implementation, with a worked candidate-selection example. BERT and learned encoders remain drafted continuation material and may overflow beyond this session. Deck - Recording | - |
| - | 21-23 Sep | No class - Chuseok week | - |
| 7 | Mon 28 Sep | LLM embeddings and reranking. LLM-based embedding models, ColBERT and MaxSim, reciprocal rank fusion, cross-encoders and language models as rerankers. Deck - Recording | - |
| 8 | Wed 30 Sep | Discussion. BEIR. Live evaluation case | Preparation, Tue 18:00 |
| - | Mon 05 Oct | No class - substitute holiday | - |
| 9 | Wed 07 Oct | Hard queries: reasoning and ambiguity. Reasoning-intensive queries, BRIGHT task design and one ReasonIR training example; ambiguity as underspecification, clarification and abstention. Deck - Recording | - |
| 10 | Mon 12 Oct | Approximate nearest neighbour search and search systems. Exact versus approximate search, HNSW and the recall-latency trade-off, product quantisation and FAISS, Elasticsearch. Deck - Recording | - |
| 11 | Wed 14 Oct | Project workshop 1, 60 min. Compare FAISS setups and Elasticsearch keyword/vector/hybrid retrieval over fixed passage vectors. | In-class prediction, run and explanation |
| 12 | Mon 19 Oct | Attention as retrieval. Attention as a lookup, retrieval-head evidence and the limits of the analogy. | - |
| 13 | Wed 21 Oct | Discussion. LIMIT: On the Theoretical Limitations of Embedding-Based Retrieval | Preparation, Tue 18:00 |
| 14 | Mon 26 Oct | Memory as an architectural axis. Routing computation versus routing a lookup. Product-key memory and the Engram design. | - |
| 15 | Wed 28 Oct | Discussion. Retrieval Heads | Preparation, Tue 18:00 |
| 16 | Mon 02 Nov | Memory learned at test time. Titans and memory updates during inference. Compare learned memory with static lookup. | - |
| 17 | Wed 04 Nov | Discussion. Engram | Preparation, Tue 18:00 |
| 18 | Mon 09 Nov | Language models with an index attached. kNN-LM as a worked example, RETRO as a contrast, long context versus retrieval. | - |
| 19 | Wed 11 Nov | Project workshop 2, 60 min. Lecturer demonstrations and individual exploration of retrieval methods. | In-class experiment record; code and report draft |
| 20 | Mon 16 Nov | Retrieval-augmented generation. A RAG pipeline, chunking and reranking choices, one corpus-poisoning failure. | - |
| 21 | Wed 18 Nov | Discussion. Titans | Preparation, Tue 18:00 |
| 22 | Mon 23 Nov | Agentic retrieval and deep research. ReAct and interleaved search, search-policy rewards, citation faithfulness. | - |
| 23 | Wed 25 Nov | Agent memory: write, manage, read. The write-manage-read loop, compaction and persistent retrieval, memory operations as tools. | - |
| 24 | Mon 30 Nov | Learning what to remember. Memory management trained end to end, memory quality under a budget, parametric memory and interference. | - |
| 25 | Wed 02 Dec | Discussion. Agentic Memory | Preparation, Tue 18:00 |
| 26 | Mon 07 Dec | Evaluating agent memory. Recall benchmarks against task utility, cost accounting and the four competencies. | - |
| 27 | Wed 09 Dec | Discussion. Evaluating Memory in LLM Agents | Preparation, Tue 18:00 |
| 28 | Mon 14 Dec | Privacy, personalisation and unlearning. Deletion from an index, unlearning from weights and one personalisation or poisoning case. | - |
| 29 | Wed 16 Dec | Project workshop 3, 60 min. Lecturer-led submission highlights and held-out comparisons. | Final code and report freeze 09 Dec, 23:59 |
Lectures are recorded, and each recording appears in the row of the session it records once it is up. Discussion sessions and project workshops are not recorded.
Amended 1 September 2026. Until today this said that project workshops are recorded too. They are where you present your own unfinished work, and that is owed the same absence of a tape as a Wednesday.
5. The individual paper-search project
Build an individual retrieval method over a fixed corpus from the reading list. The course provides extracted text, baselines, execution tools and a shared comparison interface. You submit Python code and an explanation of at most 600 words, with run references and an AI/resource declaration.
- 14 October: guided FAISS comparisons over fixed passage vectors, plus Elasticsearch keyword/vector, hybrid and metadata-filter experiments.
- 11 November: explore your own approach after lecturer demonstrations; save code, experiment evidence and a report draft in class.
- 16 December: lecturer-led highlights and held-out comparisons. No separate practical assessment or required individual presentation.
All three run 09:00-10:00 KST. Start directly in the 14 October workshop. Required work fits within class. Revisions are optional until 9 December, 23:59 KST. Without revisions, the saved November work is final.
You may change encoders, query processing, chunks, paper representations, retrieval or ranking within the shared offline CPU budget. The corpus is public to the class; final test queries are private. Numerical limits and starter access instructions will be supplied before exploration. No paid account is required.
See the project instructions for the submission interface, exploration boundaries, marking criteria, AI rules and absence arrangements.
Amended 27 September 2026. This replaces the class-note collection, group project and judging quotas. The first workshop moves from 7 to 14 October; 7 October is a lecture. The three project meetings now last 60 minutes.
6. Assessment
Amended 16 September 2026. The preparation note and individual closing paragraph are graded. Preparation assessment covers the argument, reasoning and evidence, and remaining question. Selected follow-up responses are assessed under Contribution to discussion, 10%. The oral group outcome, AI disclosure and private conduct questionnaire have no separate mark.
Both tutors mark each written piece independently, and the mark is the average of their marks. The earlier review and exit-question formats are superseded. Detailed written-work criteria and the weights of the preparation note and closing paragraph will be confirmed separately. The contribution component remains 10% and the project 25%.
Your written submissions are visible to you and the teaching team. Other students cannot read them. Peer conduct reports do not automatically affect marks.
There is no exam. In-lecture quizzes support learning and carry no marks.
The two in-class project activities carry 2 and 3 marks respectively. An authorised absence receives an equivalent opportunity. The former judging completion gate and automatic grade-band penalty are withdrawn.
Contribution to discussion, 10%
From the BEIR discussion on 30 September, the lecturer selects one or two students per group for follow-up questions within 09:30-09:40 KST. The same number is selected from each group, with one per group as the default. Selection is random among students with the fewest previous calls. Every group is asked at least once each session. Both tutors score each response, and the mark is the average of their scores. Everyone should prepare, but not everyone speaks to the whole class each session. The aim is a similar number of turns across the course.
| Score | For |
|---|---|
| 2 | Answers the question with a reason, and addresses the follow-up or acknowledges a limitation |
| 1 | Answers, but repeats the conclusion or preparation note without addressing the challenge |
| 0 | Cannot answer the question |
Marks assess reasoning, not fluency, confidence or agreement with the lecturer. The mean of recorded scores is multiplied by five for the mark out of ten. Students not selected or not reached receive no mark from that draw. No recorded response means unassessed, not zero.
See the contribution policy for availability, the call tally and the dated amendment. This replaces the earlier plan for about three turns across challenge and synthesis.
Project, 25%
The individual project is marked out of 25.
| Evidence | Marks |
|---|---|
| 14 October guided prediction, run and explanation | 2 |
| 11 November in-class experiment record | 3 |
| Final method and justified implementation | 6 |
| Final evaluation and diagnosis | 10 |
| Reproducible final execution | 4 |
The final solution and report carry 20 of 25 marks. Leaderboard position, complexity and selection for the December highlights carry no marks. A negative experimental result can receive full credit. There is no peer multiplier or separate December practical. See the project marking criteria.
7. Generative AI policy
AI may be used for preparation if declared. It is optional. Give the prompts, model outputs, what was wrong or missing, and how your writing differs. If you found no error, say what you checked. Write “None” if you used no AI.
Project-specific AI rules are in the project instructions.
AI is prohibited during the discussion session, including the individual closing paragraph. The course tutor is unavailable to students on the website and Slack during that session. Use English throughout.
The end-of-session questionnaire lets group members report direct observations of AI use or language concerns to the teaching team privately. Reports do not automatically change marks. See the discussion policy.
This 16 September amendment replaces the mandatory model check and the former AI-dependent review score rules. Cite the source for factual claims and reported numbers. Claims about your own experiments must match your actual runs.
8. Reading list
The seven discussion anchors and their role-specific supporting-paper requirements remain assigned. The lecture bibliography is a reference catalogue: only the core topics in the schedule are planned for class. The full reading list is public and gives every paper by session, with a publisher link where one exists. Web search, query likelihood and feedback, detailed ANN and multi-vector serving methods, and the additional memory and agent architectures are optional extensions. They are not catch-up homework.
Amended 26 September 2026. The list had eight anchors. The Neural Hype, ColBERT and AmbigQA discussions were withdrawn so the memory and agent material could carry more weight; ColBERT and AmbigQA are taught in lectures instead. Retrieval heads, Titans and Agentic Memory were added, and Agentic Memory replaces Search-R1, which is now a supporting paper. Later the same day, LIMIT replaced BRIGHT, which is now taught in the hard-queries lecture.
| # | Session | Paper |
|---|---|---|
| 1 | 8 (30 Sep) | Thakur et al. (2021), BEIR (NeurIPS Datasets and Benchmarks) |
| 2 | 13 (21 Oct) | Weller, Boratko, Naim & Lee (2026), On the Theoretical Limitations of Embedding-Based Retrieval (ICLR) |
| 3 | 15 (28 Oct) | Wu, Wang, Xiao, Peng & Fu (2025), Retrieval Head Mechanistically Explains Long-Context Factuality (ICLR) |
| 4 | 17 (4 Nov) | Cheng et al. (2026), Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models - the Engram paper |
| 5 | 21 (18 Nov) | Behrouz, Zhong & Mirrokni (2025), Titans: Learning to Memorize at Test Time (NeurIPS) |
| 6 | 25 (2 Dec) | Yu et al. (2026), Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents (ACL) |
| 7 | 27 (9 Dec) | Hu, Wang & McAuley (2026), Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR) |
Every anchor after BEIR is from 2025 or 2026, and that is the point. An earlier draft of this course required every assigned paper to be two to six years old, so that a real contradicting paper, a real predecessor and a reproduction all existed. That rule produces a tidy syllabus about a field that has moved on. It is gone. Sessions 17, 25 and 27 assign work whose lineage is still being written, and part of the exercise is judging a paper without the safety of a literature that has already judged it.
Four notes on the choices.
- Session 8 opens the discussion series, four weeks in. BEIR introduces evaluation before the individual paper-search project.
- Session 13 asks what one vector per document can represent at all. LIMIT proves a bound on the document sets any query can return, then builds simple queries that the embedders from the lecture fail, while BM25 does well. It sits after BEIR for a reason the class should find itself.
- Sessions 15 to 25 follow retrieval inward and then outward: attention read as lookup, a lookup built into the architecture, a memory written during inference, and a memory an agent manages through its own actions.
- Session 27 closes the loop to the evaluation lecture with 2026 material. A benchmark finds that no memory method masters all four competencies it defines; the adversary’s paper then finds that swapping only the embedding model flips which method wins.
Choosing your supporting paper: the rules
Amended 16 September 2026. Choose one supporting paper for your assigned preparation role and name it, with a link, in your preparation note. Consult its relevant sections; no full second-paper review is required.
- Choose it by the preparation deadline, Tuesday 18:00. No prior approval is required.
- It must be checkable through a stable URL, DOI or arXiv link.
- Choose outside the seven discussion anchors. Do not reuse your own earlier choice.
- Different students may choose the same paper. No coordination is required.
- Research papers, book chapters and technical reports count. Applier and Referee may also use documentation, code or public review threads where these supply the evidence.
- Explain which claim it accepts, challenges, qualifies or extends. A predecessor may explain what the anchor changed. Translate any passage you rely on into English.
- Recommendations are optional. Using one carries no penalty.
Roles guide your preparation; they do not fix the view you must defend in class. The discussion policy sets the current form and assessment scope. The older role checklists and automatic score ceilings below are superseded. Detailed marking criteria and weights will be confirmed separately.
Recommended supporting papers
These are recommendations, not assignments. Each is a safe choice that passes its role’s test, and taking one carries no advantage and no penalty. The rest are published with the roster two weeks before each session.
| Session | Role | Paper |
|---|---|---|
| 8 | Adversary | Ni et al. (2022), Large Dual Encoders Are Generalizable Retrievers |
| 8 | Historian | Nguyen et al. (2016), MS MARCO |
| 8 | Replicator | Kamalloo, Thakur, Lin et al. (2024), Resources for Brewing BEIR |
| 8 | Applier | Lin et al. (2021), Pyserini |
| 13 | Adversary | Bangachev, Bresler & Kogan (2026), Is Dimensionality a Barrier for Retrieval Models? |
| 13 | Referee | The paper’s ICLR 2026 review thread, on OpenReview |
| 15 | Referee | The paper’s ICLR 2025 review thread, on OpenReview |
| 17 | Historian | Sukhbaatar et al. (2015), End-To-End Memory Networks (NeurIPS) |
| 17 | Referee | No thread exists for this paper. You write the first review, from the reviewer guidelines alone |
| 17 | Applier | Berges et al. (2024), Memory Layers at Scale |
| 21 | Referee | Titans’ NeurIPS 2025 review thread, on OpenReview |
| 21 | Applier | Eyuboglu et al. (2025), Cartridges |
| 25 | Historian | Jin et al. (2025), Search-R1 (COLM) |
| 25 | Referee | ACL reviews are not public. You write the first review, from the reviewer guidelines alone |
| 27 | Adversary | Wang (2026), MemDelta: Controlled Baselines and Hidden Confounds in Agent Memory Evaluation |
| 27 | Referee | The paper’s ICLR 2026 review thread, on OpenReview |
Everyone holding the Referee role also reads the NeurIPS reviewer guidelines. Two pages, and it is what the role is being asked to do.
Optional evaluation reading and reference texts
These are optional references on relevance judgements and incomplete evaluation. No additional paper pre-reading is required for the project.
- Voorhees (2000), Variations in relevance judgments and the measurement of retrieval effectiveness (IP&M). TREC’s own finding that judges disagree substantially yet system rankings stay stable.
- Zobel (1998), How Reliable Are the Results of Large-Scale IR Experiments? (SIGIR).
- Buckley & Voorhees (2004), Retrieval Evaluation with Incomplete Information (SIGIR).
Reference text. Manning, Raghavan & Schütze, Introduction to Information Retrieval (Cambridge University Press, 2008). Free online. The classical retrieval chapters cover sessions 2, 3, 5 and 7.
Further reading by topic
Where knowledge sits, the classical lineage. Weston, Chopra & Bordes (2014) Memory Networks; Graves et al. (2014) Neural Turing Machines; Graves et al. (2016) DNC (Nature); Lample et al. (2019) Large Memory Layers with Product Keys; Shazeer et al. (2017) Outrageously Large Neural Networks; Fedus, Zoph & Shazeer (2022) Switch Transformers (JMLR); Jiang et al. (2024) Mixtral of Experts; He (2024) Mixture of a Million Experts.
Where knowledge sits, 2025 and 2026. Berges et al. (2024) Memory Layers at Scale; Cheng et al. (2026) Conditional Memory via Scalable Lookup - the Engram paper, and session 21’s anchor; Behrouz, Zhong & Mirrokni (2025) Titans (NeurIPS); Behrouz et al. (2025) Nested Learning (NeurIPS); Eyuboglu et al. (2025) Cartridges; Lin et al. (2025) Continual Learning via Sparse Memory Finetuning, which is the counter-example to this course’s own argument that weights cannot forget.
Attention as retrieval. Wu et al. (2024) Retrieval Head Mechanistically Explains Long-Context Factuality, the empirical version of an analogy this course used to assert; Zhang & Bottou (2025) Memory Mosaics at scale (NeurIPS).
Long context versus retrieval. Liu et al. (2023) Lost in the Middle (TACL); Li et al. (2024) Retrieval Augmented Generation or Long-Context LLMs? (EMNLP).
Dense retrieval, and one space for two jobs. Reimers & Gurevych (2019) Sentence-BERT (EMNLP); Karpukhin et al. (2020) DPR (EMNLP); Ni et al. (2022) Large Dual Encoders Are Generalizable Retrievers (EMNLP); Muennighoff et al. (2024) Generative Representational Instruction Tuning; Zhang et al. (2024) OneGen (EMNLP Findings); Lee et al. (2025) Gemini Embedding; Zhang et al. (2025) Qwen3 Embedding; Enevoldsen et al. (2025) MMTEB (ICLR), the benchmark those models are ranked on, and the paper that shows what removing the training sets from it does to the ranking.
Retrieval that needs reasoning. Su et al. (2025) BRIGHT (ICLR); Shao et al. (2025) ReasonIR.
What one vector can represent. Weller, Boratko, Naim & Lee (2026) On the Theoretical Limitations of Embedding-Based Retrieval (ICLR), the LIMIT discussion anchor; Bangachev, Bresler & Kogan (2026) Is Dimensionality a Barrier for Retrieval Models?
Multi-vector retrieval, and making it servable. Khattab & Zaharia (2020) ColBERT (SIGIR); Santhanam et al. (2022) ColBERTv2 (NAACL) and PLAID (CIKM); Dhulipala et al. (2024) MUVERA (NeurIPS); Veneroso et al. (2025) CRISP; Faysse et al. (2025) ColPali (ICLR).
Nearest neighbour search and systems. Malkov & Yashunin (2018) HNSW (TPAMI); Jégou, Douze & Schmid (2011) Product Quantization; Johnson, Douze & Jégou (2019) Billion-scale similarity search with GPUs.
Learning to rank and click bias. Burges (2010) From RankNet to LambdaRank to LambdaMART; Joachims et al. (2017) Unbiased Learning-to-Rank with Biased Feedback (WSDM); Craswell et al. (2008) An Experimental Comparison of Click Position-Bias Models (WSDM).
Retrieval-augmented generation. Lewis et al. (2020) RAG; Borgeaud et al. (2022) RETRO; Khandelwal et al. (2020) kNN-LM (ICLR); GraphRAG (2024); Gutiérrez et al. (2025) From RAG to Memory (ICML), which is HippoRAG 2.
Agentic retrieval and deep research. Yao et al. (2023) ReAct; Trivedi et al. (2023) IRCoT (ACL); Asai et al. (2023) Self-RAG; Jin et al. (2025) Search-R1 (COLM), session 25’s anchor; Chen et al. (2025) BrowseComp-Plus, which re-evaluates that literature on a fixed corpus and finds much smaller numbers; Liu, Zhang & Liang (2023) Evaluating Verifiability in Generative Search Engines (EMNLP Findings), on whether citations support the claims attached to them.
Agent memory, and whether it can be measured. Hu, Wang & McAuley (2026) Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR), session 27’s anchor; Wang (2026) MemDelta, on hidden confounds in that evaluation; Anthropic, Effective context engineering for AI agents, read critically as a primary source.
Semantic search in practice. Cormack, Clarke & Buettcher (2009) Reciprocal Rank Fusion (SIGIR); Formal et al. (2021) SPLADE; Nogueira et al., doc2query.
Adversarial retrieval. Greshake et al. (2023) Not what you’ve signed up for (AISec); Zou et al. (2024) PoisonedRAG.
Commercial search. Shah & Bender (2022) Situating Search (CHIIR), the strongest case against replacing search with language models, and worth reading precisely because this course argues the other way; Google’s How Search Works, read critically as a primary source.
Scholarly search and literature synthesis. Lo et al. (2020) S2ORC (ACL); Swanson (1986) on literature-based discovery; Wang et al. (2024) SciMON (ACL); Asai et al. (2024) OpenScholar; Skarlinski et al. (2024) PaperQA2. This block lost its own lecture in the 2026 revision and is now read alongside session 24, because deep research and literature synthesis are the same problem.
9. Logistics
Language: English
Lecturer: Seong Joon Oh
Tutors: 김현재 and 이태훈 (Taehoon Lee).
When: Mondays and Wednesdays, 09:00-10:00 KST, including project sessions.
Where: Online (Zoom), for every session. One link carries the whole term; it is on the course home page once you are signed in, and it was emailed before the first session. Cameras are required in discussion groups and optional in plenary.
The course registration system lists a classroom at 양재산학캠퍼스 against both slots. That room is not used. Do not go there.
Enrolment: Capped at 40.
Email: coallaoh@stai-lab.org for submissions, questions, and feedback.
Slack: The course workspace carries announcements, the weekly roster, and questions. Join here.
Discussion groups. You will be in a fixed group of four or five for discussion sessions. Propose your own group on the first-day survey by Monday 14 September; remaining places are filled at random after that.