an interesting conversation on X about the way it is becoming difficult to maintain up with recent research papers due to their ever-increasing quantity. Truthfully, it’s a general consensus that it’s unimaginable to maintain up with all of the research that’s currently happening within the AI space, and if we should not in a position to sustain, we’re then missing out on quite a lot of necessary information. The essential crux of the conversation was: who’re we writing for if humans can’t read it, and if LLMs are those actually reading the papers, what’s the best format for them?
This had me considering and it jogged my memory of an article I wrote back in 2021 on the tools I used to read research papers effectively and the way I read papers back then. That was the pre-ChatGPT era, and I realised how much paper reading has modified for me, since then.
So I’m sharing how I read research papers today, each manually and with AI assistance. My hope is that for those who are also getting overwhelmed by the pace, a few of these ideas or tools might assist you construct a flow that works for you. , but I can not less than share what has worked for me up to now.
The Manual way — three-pass method style
There was a time when all of the reading was manual and we used to either print papers and skim them or achieve this via an e-reader. During that point I used to be introduced to a paper by S. Keshav on the three-pass method. I’m sure it’s essential to have also come across it. It’s an easy yet elegant way of reading a paper by breaking the method into three steps.
As shown within the figure above, the three-pass method helps you to control how deep you must go based in your purpose and the time you’ve. Here’s what each pass involves:
- The primary pass gives a fast bird’s-eye view. You scan the paper to know its essential idea and check if it’s relevant. The goal is to reply the 5 Cs at the top of your reading : the category of the paper, its contribution, whether the assumptions are correct, the clarity of the writing and the context of the work. This shouldn’t take greater than 5–10 minutes.
- The second pass can take as much as an hour and goes a bit deeper. You may make notes and comments, but skip the proofs for now. You primarily must give attention to the figures and graphs and take a look at to see how the ideas connect.
- The third and final pass takes time. By now the paper is relevant, so that is the stage where you read it fastidiously. You must have the option to trace the complete argument, understand the steps and mentally recreate the work. This can be where you query the assumptions and check if the ideas delay.
Even today, as much as possible, I attempt to begin with the three-pass method. I actually have found it useful not only for research papers but in addition for long technical blogs and articles.
The Chatbot summary way — vanilla style
Today, it’s easy to drop a paper into an LLM-powered chatbot and ask for a fast summary. Nothing improper in that, but I feel most AI summaries are quick and at times flatten the ideas.
But I actually have found few prompts that work higher than the vanilla “summarise this paper” input. As an illustration, you may ask the LLM to output the summary in a three-pass style, the identical method we discussed within the previous section which provides a significantly better output.
Give me a three-pass style take a look at this paper.
Pass 1: a fast skim of what the paper is about.
Pass 2: the essential ideas and why they matter.
Pass 3: the deeper details I should concentrate to.One other prompt that works well is a straightforward problem–idea–evidence style:
Tell me:
• what problem the paper tries to unravel
• the essential idea they use
• how they support it
• what the outcomes mean.Or if I would like to ascertain how a paper compares with past work, I can ask:
Give me the essential idea of the paper and likewise indicate its limits or things
to watch out aboutYou may all the time proceed the chat and ask for more details if the primary answer feels light. However the essential issue for me remains to be the identical: it’s essential switch between tabs to have a look at the paper after which compare the reason and each sit elsewhere. For me, that constant back-and-forth becomes a degree of friction. There must be a greater way which keeps each the source and AI assistance on the identical canvas and this takes us to the subsequent part.
The specialised tools way — UI matters
So I got down to explore tools that provide LLM-assistance yet offer a greater UI and a smoother reading experience. Listed below are three that I’ve used personally. This isn’t an exhaustive list, just those that, in my experience, work well without replacing the core reading experience. I’ll also indicate out the features that I like essentially the most for each tool.
1. alphaXiv
AlphaXiv is the tool I’ve been using for a very long time since it has many useful things built right into the platform. It’s easy to achieve a paper here, either through their feed or by taking any arXiv link and replacing arxiv with alphaxiv. You get a clean interface and a bunch of AI-assisted tools that sit right on top of the paper. There’s a well-recognized chat window but aside from which you can highlight any a part of the paper and ask an issue right there. You may also pull in context from other papers using the @ feature. If you must go deeper, it shows related papers, the GitHub code, how others cite the work and small literature notes across the topic, as well. There’s an AI audio lecture feature too, but I don’t use it often.
My favourite part is the blog-style mode. It gives me an easy, readable version of the paper that helps me determine if I should do a full deep read or not. It keeps the figures and structure in place, almost like how I’d turn a paper right into a blog.
- Learn how to Try: Replace with in any arXiv link, or open it directly from their site at alphaxiv.org.
2. Papiers
How do you discover recent papers? For me it’s through just a few newsletters, but more often than not it’s from some outstanding X accounts. Nevertheless, the issue is that there are lots of such accounts and so there’s quite a lot of noise and signal has grow to be harder to follow. Papiers aggregates conversations a few paper and other papers related to it into one place, making the invention a part of the reading flow itself.
Papiers is a reasonably recent tool but already has some great features. As an illustration, along with getting conversations in regards to the paper, you may get a Wiki-style view in two formats — technical and accessible so you may select the format based in your comfort level with the subject. There’s also a Lineage view that shows the paper’s parents and kids, so you may see what shaped the work and what got here after it. And there’s also a mind map feature (think NotebookLM) that’s pretty neat.
I desired to indicate here that the tool did give me error for some papers, or the X feed was missing for just a few. It did work for the outstanding papers though. I looked around and located in a X thread that papers currently get indexed on demand, so I suppose that explains it. But it surely’s a brand new tool and I actually just like the offerings, so I’m sure this part will improve over time.
- Learn how to Try : Replace with papiers in any arXiv link, or open it directly from their site at papiers.
3. Lumi
Lumi is an open-source tool from the People + AI Research group at Google and as with quite a lot of their work, it comes with a surprising and thoughtful UI. Lumi highlights the important thing parts of the paper and places short summaries within the side margin, so you usually get to read the unique paper together with AI generated sumamry. You may also click on any reference and it takes you straight to the precise sentence within the paper. The standout feature of Lumi is that it not only explains the text but you may also select a picture and ask Lumi to elucidate it as well.
The one downside is that it currently works for arXiv papers under a Creative Commons license, but I’d like to see it expand to cover all of arXiv and perhaps even allow uploading PDFs of other papers.
Other tools price a mention
While I mostly use the above mentioned tools, there are just a few others that I’ve definitely crossed paths with, and I’d encourage you to try them out in the event that they suit your flow like: They didn’t grow to be my essential decisions, but they do have some good ideas and might work well for you depending in your reading style.
- OpenRead is a terrific option for reading papers in addition to doing literature survey. It has some great add-ons like comparing papers, paper graphs to indicate connected papers and a paper espresso feature that offers a concise one pager summary of the paper.
Something to notice here is that OpenRead is a paid tool but does include a freemium version.
- SciSpace is a really versatile tool and along with with the ability to chat with a paper, you may do semantic literature reviews, go deep into research, write papers and even create visualisations to your work. There are a lot of other things it offers, which you’ll explore of their suite. Like OpenRead, additionally it is a paid tool with limited features available within the free tier.
- Each day Papers by HuggingFace is great option for those who want to see trending papers to see trending papers. One other nice touch about his is you may immediately see the models, datasets and spaces on HuggingFace citing a selected paper (in the event that they exist) and likewise chat with the authors.
Conclusion
A lot of the reading that I do is a component of the literature review for my blog, and it’s a combination of the three strategies that I discussed above. I still like going through papers manually, but when I would like to go further, see connected papers or understand something in additional detail, the three tools I discussed help me rather a lot. I’m aware that there are lots of more AI-assisted tools for reading papers, but similar to the phrase , I prefer to keep on with just a few and never jump between favourites unless there’s a very standout feature.
