Research papers are read on the subject of LLMS

Interesting discussion on X about how it's getting harder to keep up with new research papers because of their increasing size. Honestly, it's a general agreement that it's impossible to keep up with all the research that's happening right now in the AI ​​space, and if we can't keep up, we've missed out on a lot of valuable information. The big crux of the discussion is: Who are we writing to if people can't read it, and if llms are the ones actually reading the papers, what format is right for them?

This got me thinking and reminded me of an article I wrote back in 2021 on the tools I was learning to read research papers effectively and how I read papers back then. That was an early period, and I realized how much paper reading has changed for me, since then.

So I'm sharing how I read research papers today, manually and with the help of AI. My hope is that if you're also overwhelmed by speed, some of these ideas or tools can help you create a flow that works for you. I don't really have an answer to what the ideal paper format should look like for the LLM ERAbut at least come in and share what has worked for me so far.

Manual method – a three-pass style

There was a time when all reading was a notebook and we worked or printed papers and read them or did so with an e-student. At that time I was introduced to the paper by S. Keshav to A three-pass road. I'm sure you must have found it again. It is an easy and simple way to study the paper by breaking the process into three steps.

As shown in the figure above, the three-way approach allows you to control how far you want to go depending on your purpose and the time you have. Here's a throwdown for each including:

1. First pass gives a quick glance. You scan the paper to understand its main idea and assess its relevance. The goal is to respond 5 cs At the end of your reading: paper section, this presentationthat consideration is something – appropriatethe to explain of writing and totality of work. This should not take more than 5-10 minutes.
2. The second pass can take an hour and is a little more relaxing. You can make notes and comments, but skip the proof now. You mainly need to focus on the structures and graphs and try to see how the ideas connect.
3. The third and final pass takes time. By now you know the paper works, so this is the section where you read it carefully. You should be able to follow a complete argument, understand the steps and redo the work. This is also where you question assumptions and test whether the assumptions hold.

Even today, as much as possible, I try to start with more than three. I have found it useful not only for research papers but also for long technical blogs and articles.

Chatbot Shortcut – Vanilla Style

Today, it's easy to drop a piece of paper into a powerful LLM Chatbot and ask for a quick summary. There's nothing wrong with that, but I feel a lot of AI summaries are quick and sometimes soften the ideas.

But I've found a few stimulants that work better than vanilla”summarize this paper“Input. For example, you can ask LLM to output a summary in style 3, the same method we discussed in the previous section that gives a better result.

Give me a three-pass style look at this paper.
Pass 1: a quick skim of what the paper is about.
Pass 2: the main ideas and why they matter.
Pass 3: the deeper details I should pay attention to.

Another quick that works well is a simple problem-idea – style proof:

Tell me:
• what problem the paper tries to solve
• the main idea they use
• how they support it
• what the results mean.

Or if I want to check how a paper compares to previous work, I can ask:

Give me the main idea of the paper and also point out its limits or things 
to be careful about

You can always continue the conversation and ask for more details if the initial response is sound. But the main problem for me is still there: you need to switch between tabs to look at the paper and compare the description and both stay in different places. For me, that back and forth has always been a point of contention. There has to be a better way that keeps source and AI help on the same canvas and this takes us to the next part.

Special Approach to Tools – UI Matters

So I started exploring tools that provide LLM support but offer a better UI and smoother learning experience. Here are three that I have personally used. This is not an exhaustive list, this is just what, in my experience, works well without giving back the basic learning experience. I will also highlight the features I like most about each tool.

1. Alphaxiv

Alphaaxiv is a tool I've been using for a long time because it has so many useful features built right into the platform. It's easy to access the paper here, either through their feed or by taking any arxiv link and retrieving the rest arxiv and alphaxiv. You get a clean interface and a bunch of ai-assisted tools that sit on top of the paper. There is a standard chat window but other than that you can do it heat it up Any part of the paper then ask a question there. You can also pull context from other documents using @ Feature. If you want to dig deeper, it shows related papers, GitHub code, how others are citing the work and smaller articles that delve into the topic, too. There is an AI Audio speech feature as well, but I don't use it very often.

My favorite part is Blog style mode. It gives me a simple, readable form of paper that helps me decide whether or not to do a deep dive. It keeps the math and structure in place, almost like how I would turn a paper into a blog.

• How to try: Put in place arxiv and alphaxiv from any arxiv link, or open it directly from their site at Alphaaxiv.org.

2. Papers

How do you get new papers? For me it goes through a few stories, but most of the time it comes from some top X accounts. However, the problem is that there are many such accounts and therefore there is a lot of noise and the signal has become difficult to follow. Papers brings together discussions about a paper and other related papers in one place, making the discovery part of the study flow.

Papiers is a fairly new tool but it already has some great features. For example, in addition to receiving interviews on paper, you can receive Wiki-Style View in two formats – technical and accessible to choose the format according to your comfort level with the topic. There is also Edible storage space A view that shows the parents and children of the paper, to see what is watching this activity and what followed after it. And there's also a mind map feature (think notebooklm) which is neat.

I wanted to point out here that this tool gave me this paper not found An error in some papers, or the X feed was missing in a few. Outstanding papers worked. I've looked around and found on the X thread the papers currently have a clue about the need, so I guess that explains it. But it's a new tool and I really like the contributions, so I'm sure this feature will improve over time.

• How to try: Put in place arxiv For the papers at any arxiv link, or open them directly on their site in error.

3. Dry

Lumi is an open source tool from the human + AI research team at Google and like most of their work, it comes with a stunning and thoughtful UI. Lumi highlights important parts of the paper and puts short summaries in the margin, so you always get to read the original paper and the AI-generated sumamry. You can also click on any index and it takes you straight to the exact author on the paper. The feature of Lumi stand is that it doesn't just explain the text but you can just select a picture and ask Lumi to explain as well.

The only downside is that it currently works on artxiv papers under a Creative Commons license, but I'd like to see it expand to cover all of arxiv and maybe allow uploading PDFs of other papers.

Some tools are worth mentioning

While I mostly use the tools mentioned above, there are a few others that have crossed paths, and I would encourage you to try them out if they find your best flow, but they have great ideas and may work well for you depending on your learning style.

• Open it it's a great way to read papers and do a literature review. It has great additions like compare papers, Paper graphs to show linked papers and a Paper Espresso A feature that provides a short pager summary of the paper.

Something to note here is that OpenFree is a paid tool but it comes with a freemium version.

• Scispace it is a very versatile tool and in addition to being able to discuss the paper, you can do a semantic literature review, delve into research, write papers and create a view of your work. There are many other things they offer, which you can check out on their site. Like OpenFree, it is also a paid tool with limited features available in the free TIER.
• Papers every day With HuGingGraface is a good way if you want to see trending papers to see trending papers. Another nice touch is that it is possible to quickly see the models, datasets and spaces on huggainface that present a specific paper (if they exist) and discuss with the authors.

Lasting

Most of the reading I do is part of my book review blog, and it's a mix of the three techniques I mentioned above. I still like to go to the papers manually, but when I want to go further, look at linked papers or understand something in detail, the three tools I mentioned help me a lot. I know there are many tools that help AIs to read papers, but as a word Too many cooks spoil the brothI like to stick to a few and not jump between favorites unless there is a really stand out feature.