An Efficient Way to Read Papers with LLMs

1. Start with an LLM-generated summary

There are many ways to chat with a PDF. The tool I use most often these days is Claude, although SciSpace, ChatPDF, ChatDOC, DanteAI, and many others are also available. In my experience, ChatGPT processes PDFs in chunks using a RAG-style approach rather than using the entire file at once, which makes it difficult to obtain a proper summary of the paper as a whole.

Because I am usually interested in the specific research methodology, I ask for an explanation of the exact experimental settings and explicitly request that numerical details not be omitted. After this first pass, I understand the overall paper and can read the main text with a purpose: I know which parts I need to examine more closely.

2. Read the paper and take Obsidian notes

Rather than reading every word from the beginning without a plan, I first build a broad map of the paper in my mind. This lets me read strategically, knowing which details belong in which parts of that map. As I read, I organize the points I want to remember as an itemized list in Obsidian.

I usually structure a note in the following way. STRI is inserted consistently at the end through a template.

• A two-line summary
• A running summary of the paper while reading
• Concept notes for important keywords that appear along the way
• Source: usually a link to the paper
• Topic: one or two higher-level concepts used for classification, so that clicking a topic note reveals papers in the same category
• Related: links to other paper notes or concept notes related to this paper
• Ideas: ideas that use the paper’s content, or limitations and implications not mentioned by the authors. This is a form of thinking practice.

3. Ask an LLM to review the notes

I give the LLM my notes and ask it to point out anything I may have misunderstood or any important points I may have missed. I currently use Claude, and I find its feedback highly satisfactory. When it identifies something I missed, I return to the main text and correct my notes.

It is also possible to use a question-and-answer format to check parts that still seem uncertain.

4. Revisit old notes regularly

This is part of spaced retrieval for strengthening long-term memory. One way to do this is to read other notes connected through the “Related” or “Topic” fields.

The index in my own long-term memory needs to be organized properly. Then, when I write a paper later, I can immediately recall that I have read a relevant reference.

Why I Write the Notes Myself

Simply generating and storing paper summaries has little value. I need to go through the process of writing in my own words.

Knowledge that I have read and understood is implicit knowledge; it becomes explicit knowledge only when I express it in writing. As I try to articulate it explicitly, I often discover parts that I thought I understood but had actually passed over.

LLM feedback and follow-up questions can help fill in incorrect or incomplete parts. The act of expressing something in explicit language is itself a form of encoding, helping move it from short-term memory into long-term memory.

If I then strengthen that long-term memory through spaced retrieval by revisiting old notes, it is icing on the cake. The reference notes accumulated in this way also become excellent material for writing papers later. I will write about my approach to writing on another occasion.