Paper
study group.

Why this group exists

Reading one paper in depth teaches you one paper. Listening to five colleagues explain theirs teaches you six. The Paper Study Group is built on that arithmetic: every two weeks a panel of six people each present, in ten minutes, a paper they have read and truly understood, so everyone leaves with a working grasp of six ideas instead of one.

The topics are generative AI, large language models, machine learning and AI more broadly. We lean towards what is modern, and towards what could matter for our work in banking: improving the customer journey, choosing the next best action, cross-selling. Relevance is welcome, never required. A good idea from outside our field is often the one that changes how we work.

What a session looks like

A session lasts 70 minutes: six presenters, ten minutes each, and ten minutes to keep the changeovers calm.

Time	What happens
0:00	Welcome, running order, the rules in one sentence (2 min)
0:02 – 1:02	Six slots of 10 minutes: 8 minutes of talk, 2 minutes of questions
between slots	A one-minute changeover, five in all, so the next presenter is ready
1:07	Wrap-up: one takeaway each, what we want to try, next session (3 min)

The ten minutes are a hard limit, enforced by a visible timer. A short talk forces a clear one. If a paper deserves more, that is a sign to schedule a follow-up, not to run over.

Who takes part

The panel is five or six presenters. Everyone is welcome to join as a spectator: no paper, no preparation, just curiosity. Spectators are part of the session, not an audience of ghosts. They are asked to bring one question for at least one of the talks, and anyone who wants to move from the seats to the panel can do so at the next session.

Presenting rotates, so that over a few sessions everyone who wants to has had a turn and nobody is asked to present every time.

How to give an eight-minute talk

The aim is understanding, not coverage. A good talk answers six questions, roughly in this order:

1. What problem does the paper solve, and why is it hard? (1 min)
2. What is the core idea, in one sentence and one picture? (2 min)
3. How does it work, at the level you could redraw it from memory? (2 min)
4. What is the evidence, and do you believe it? (1 min)
5. Where does it fail or stop short? (1 min)
6. What would it take to try this in our own setting? (1 min)

Use at most five slides, or none. Nobody reads from the slides. Papers are public, so the talks never need internal data or numbers.

What counts as having read it in depth

You are ready to present when you can do all of these without the paper in front of you: explain the problem to a colleague who has not heard of it, redraw the method, name the baseline it beats and by how much, state one real limitation, and say what you would need in order to try it here. This usually takes four to eight hours of reading over a week or two, and it is the point of the whole exercise.

Every presenter also writes a one-page paper card with the same questions, which is published with the session so the work outlives the talk.

Choosing papers

Each presenter chooses their own paper, from the reading shortlist or from anywhere else. The shortlist is fed by a small scout that looks at what researchers and engineers are sharing and discussing (Hacker News, Hugging Face daily papers, arXiv) and ranks it for recency, community interest and relevance to our work. It also keeps the list varied, so one fashionable topic cannot take every slot. The latest scout results are on the radar page.

A few guidelines keep the sessions rich:

• Prefer recent work, from the last 18 months. One modern classic per session is fine.
• A paper should have real experiments or a real argument, not only a promise. A preprint is fine.
• Within one session, no more than two papers on the same topic, and at least one that is close to banking, recommendation or customer behaviour.
• A paper is presented once. The library records what has been read, so nothing repeats by accident.

The two-week rhythm

When	What
Two weeks before	The scout shortlist is refreshed; presenters pick their papers
One week before	Papers and one line on why each matters are announced
Three days before	Paper cards are in, so spectators can skim what is coming
Session day	Six talks, seventy minutes
Within two days	Slides, paper cards and takeaways are published on the session page
Afterwards	Comments and annotations stay open on each paper, and the library is updated

What is kept

Nothing is lost between sessions. This site keeps a page for each session with the running order, slides, takeaways and open questions; a page for each paper with its card, a PDF link and annotations; and a searchable library of everything read, filterable by topic and by who presented it. Annotations on the papers themselves are shared in a private annotation group, and the best ones are linked from the paper page.

Ground rules

• Be kind, and be precise. Questions are about the paper, not about the presenter.
• It is fine to say that you did not understand something. It is the most useful sentence in the room.
• Papers and ideas are public. Anything about our own data, customers or systems stays out of the talks and out of this site.
• Presenters appear by first name, and only if they agree to it.