Hacker News Reader

Comment by s1mon | Hacker News Reader

I've tried this and chatpdf with some Computer Aided Geometric Design mathematics theses and papers. These tools do not seem to "read" the math parts of the papers, only the text. It would be amazing to have something which could digest a bunch of papers and start to help me navigate them, but a lot of the answers I get are "I don't have enough context to help you". Is there a better tool or way to use LLMs with scholarly mathematics papers?

View on HN

You could just use the latex source code of the paper in the context window if you want it to be aware of the math. You can download latex source for any paper or arxiv.

If you have many papers, perhaps upload the latex source code of all of them into some vector database like pinecone and use them with LangChain for retrieval.

this is a great idea, totally doable with what is out there

Great point, I think this is important knowledge to spread. Using latex format is a sweet hack to do math based analysis for papers when using LLM frameworks. As long as you have your work in latex already or even if you can convert it to Latex. It seems to be able to understand Lyapunov stability equations, Kalman filter analysis, or Hamiltonian equations when I paste them in. There is definitely hallucinations here and there but the understanding is amazing. For instance, you can convert a Word document with Mathtype equations that use Latex and then copy it into ChatGPT and it can articulate what every line item equation is doing. With reasonable coherence. Likely its not accurate with cutting edge scientists, but pretty sweet for the engineers.

It's not always great, but try explainpaper [1].

I have nothing to do with it, but to me it's the best, they seem to put some real effort on it since day one.

[1] explainpaper.com

Comment of the day, this is why I keep checking HN daily.

you're welcome :D

It’s because embedding math equations and data requires a different paradigm

Yeah it is mediocre. I think the issue is having to first upload/read then push back data. It has not been optimized yet. It needs lot of work. I am not sure which open-source has that capability yet to run it locally. Perhaps that will be faster but I have not played with or explored that area.

Do you have any example papers i can play with?

Reply on news.ycombinator.com