Pairs, papers and discussants
Published Wednesday Nov 4. If a name is wrong, or a date clashes with an exam elsewhere, write to me before the kickoff on Friday Nov 6.
Every student appears twice on this page: once as a presenter, with their partner, and once as a discussant of a paper in another session.
| When | What |
|---|---|
| Fri Nov 6, 9:00–12:30 | Kickoff, online. From 11:30 the pairs work together |
| Tue Nov 10, 23:59 | Reading plan, half a page per pair, by email: the paper in 5 sentences, what you reproduce, what you run on the testbed, compute plan, risks |
| Wednesday before your session, 18:00 | Presenters: slides, code, results.json and demo queries
in the shared repo |
| Thursday before the session you discuss, 18:00 | Discussants: one-page note ending with two questions |
| Your session + 7 days | Presenters: short report |
What each role does
- Presenters. A 50-minute slot starting at the time given below: a 25-minute talk, a 10-minute live demo with three queries (hero, failure, wildcard), then 15 minutes of dialogue with the discussants and the room. Before the session, reproduce one key result of the paper at small scale and run the method on the shared testbed. The format and the demo protocol are in program §3; what your paper asks for is in the paper menu.
- Discussant A, the auditor. Reads the pair's code,
README and
results.jsonas closely as the paper. The one-page note ends with two questions about the implementation or the numbers. In the session, A asks first in the 8-minute question round, then moderates the 7-minute open discussion. - Discussant B, the skeptic. Reads the paper against the field: a real problem or a benchmark artefact? What does it cost to deploy? How does it relate to the spring pipeline and to another paper of the seminar? The note ends with two questions about validity and use; B asks after A and reopens the discussion if the room is quiet.
- Everyone. Before each session, one question about one of the day's papers, on the question page (Thursday 18:00).
The full briefs are in program §8, the grading sheets in grading sheets §1–3.
Every session starts at 9:00 with 30 minutes of framing by me, and the break is 11:10–11:25: come at 9:00 even when your talk is later.
Nov 13 · Representations: how embeddings are trained, and where they break
| Start | Paper | Presenters | Discussant A (auditor) | Discussant B (skeptic) |
|---|---|---|---|---|
| 9:30 | #6 Timeline talk: static embeddings are back — Model2Vec and static retrieval models (Aarsen, HF blog 2025; Model2Vec 2024) ★ | Martin Leiva · Vladimir Herrera Nativi | Juan Kenichi Sutan | Sukhjot Kaur |
| 10:20 | #2 LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders (BehnamGhader et al., COLM 2024) ★★ | Raphael Leonardi · Baptiste Pras | Oliver Wakeford | Alexandre Costa |
| 11:25 | #4 On the Theoretical Limitations of Embedding-Based Retrieval (LIMIT) (Weller et al., ICLR 2026) ★★ | Wissam Houria Amar · Zakaria Soualah Mohammed | Ran Lu | Mouhamadou Moustapha Diop |
Nov 20 · Domains: scientific papers, patents, adaptation, languages
| Start | Paper | Presenters | Discussant A (auditor) | Discussant B (skeptic) |
|---|---|---|---|---|
| 9:30 | #12 Timeline talk: Unsupervised word embeddings capture latent knowledge from materials science literature (Tshitoyan et al., Nature 2019) ★★ | Valentina Cadime · Alexandre Costa | Julie Dornat | Kseniia Ostudina |
| 10:20 | #10 GPL: Generative Pseudo Labeling for Unsupervised Domain Adaptation of Dense Retrieval (Wang et al., NAACL 2022) ★ | Francesc Serra Crespí · Marina Hornero Merino | Dominika Chojnacka | René Mourambou Diyema |
| 11:25 | #11 BGE-M3: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings (Chen et al., Findings of ACL 2024) ★★ | Mathis Lesgourgues · Oliver Wakeford | Youssef El Otmani | Matteo Lombardi |
Nov 27 · LLMs in the loop: rerankers, rewriters, reasoning
| Start | Paper | Presenters | Discussant A (auditor) | Discussant B (skeptic) |
|---|---|---|---|---|
| 9:30 | #15 HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels (Gao et al., ACL 2023) ★ | Xuanru Dai · Juan Kenichi Sutan | Marina Hornero Merino | Eman Sarfraz |
| 10:20 | #13 Is ChatGPT Good at Search? LLMs as Re-Ranking Agents (RankGPT) (Sun et al., EMNLP 2023) ★ | Said Abolhassan Razavi · Zening Wang | Vladimir Herrera Nativi | Mathis Lesgourgues |
| 11:25 | #14 Rank1: Test-Time Compute for Reranking (Weller et al., COLM 2025) ★★ | Sukhjot Kaur · Matteo Lombardi | Baptiste Pras | Hao Xiao |
Dec 4 · RAG and agents
| Start | Paper | Presenters | Discussant A (auditor) | Discussant B (skeptic) |
|---|---|---|---|---|
| 9:30 | #22 The Power of Noise: Redefining Retrieval for RAG Systems (Cuconasu et al., SIGIR 2024) ★ | Mouhamadou Moustapha Diop · Julie Dornat | Martin Leiva | Xuanru Dai |
| 10:20 | #19 Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (Asai et al., ICLR 2024) ★★★ | Dominika Chojnacka · Youssef El Otmani | Zakaria Soualah Mohammed | Mouhamed Nourou Dine Cisse |
| 11:25 | #23 Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning (Jin et al., 2025) ★★★ | Ran Lu · Hao Xiao | Valentina Cadime | Zening Wang |
Dec 11 · Trust: evaluation, attacks, bias, and documents that are not text
| Start | Paper | Presenters | Discussant A (auditor) | Discussant B (skeptic) |
|---|---|---|---|---|
| 9:30 | #27 Neural Retrievers Are Biased Towards LLM-Generated Content (Dai et al., KDD 2024) ★ | René Mourambou Diyema · Kseniia Ostudina | Said Abolhassan Razavi | Wissam Houria Amar |
| 10:20 | #28 ColPali: Efficient Document Retrieval with Vision Language Models (Faysse et al., ICLR 2025) ★★ | Mouhamed Nourou Dine Cisse · Eman Sarfraz | Raphael Leonardi | Francesc Serra Crespí |
After the two talks, from 11:25: the synthesis of the seminar, with the assembled results table.
How this was decided
Every mutual partner request is honoured. The other pairs join a student who took the spring course (or a similar course) with one who did not, as far as the numbers allow. Papers then maximise everyone's preferences, with a penalty that keeps low-ranked choices rare: 17 of the 28 students who bid have their first choice, 22 their first or second. Each discussant reviews a paper in another session, most often one they bid on themselves, and no two pairs review each other.