#Research papers

Most reading lists are undifferentiated: fifty papers, no guidance, so you read none of them. This one is sorted by what you should actually do with each paper, because that is the decision you are making.

A caution before the list: papers report results under their own conditions. A method that beat baselines on MS MARCO in 2023 may do nothing for your corpus. Read them for mechanisms and for the shape of the trade-off, not for numbers to quote.

#If you only read ten

In this order. This is roughly two weekends and it covers the mechanisms underneath everything else.

#Architecture & training

Chinchilla is the one to actually read in this section. It reframed the field — the finding that parameter count had been over-weighted relative to training tokens is why the useful small models exist at all.

#Prompting & reasoning

The last one matters more than its citation count suggests. It is the evidence behind "chain-of-thought is not an explanation", and it is the paper to cite when someone proposes showing CoT to a regulator.

#Retrieval & RAG

BEIR is the sobering one. It shows retrieval methods that dominate one benchmark falling behind BM25 on out-of-domain data — which is the honest argument for hybrid retrieval and for evaluating on your own corpus.

#Evaluation

Chouldechova (2017) and Kleinberg et al. (2016) are the fairness impossibility results. Short, mathematical, and they settle an argument that otherwise recurs forever in product meetings.

#Efficiency & serving

#Agents, tools & adaptation

LIMA is the one to read if you are about to fine-tune. It is the strongest published argument that data curation, not volume, is the work.

#How to read one of these efficiently

Most papers do not need a linear read. A method that works:

1. Abstract, then figures. The figures usually carry the contribution. If you cannot tell what the paper claims after the figures, the abstract was badly written, not you.
2. Conclusion before the method. It tells you what to look for.
3. Method section only if you need the mechanism. Often you do not.
4. Skip the related work unless you are surveying the field.
5. Read the limitations section. It is the most honest part of most papers and the fastest route to knowing whether it applies to you.
6. Check the baselines. A large improvement over a weak baseline is a weak result, and this is where over-claiming hides.

Twenty minutes per paper on that method covers far more ground than two hours of linear reading, and retains more.

#Staying current without drowning

Papers are a poor way to track a fast field — by publication, results are months old. More useful:

A discipline worth adopting: when a paper or release seems relevant, write one line in your own notes saying what you would change because of it. If you cannot write that line, you did not need the paper — and this is also the habit that turns reading into something you can talk about.

#An honest caveat

Citations here are given by title, primary author and year so you can find them. Verify details before quoting a specific number or author list — venues, author orders and reported figures are exactly the kind of detail that is easy to get subtly wrong, and being confidently wrong about a paper in an interview is worse than not having read it.

The safe form is: "the LLM.int8 paper showed that above roughly seven billion parameters, a small number of activation dimensions become extreme outliers, which is why naive INT8 breaks." That is the mechanism, it is what matters, and it does not depend on a number you half-remember.