Surveying the literature with LLMs

One of our primary usecases for using LLM-based workflows is to survey the literature. LLM-based agents are very useful for quickly obtaining an overview of what has been done in a field.

One still has to keep a few caveats in mind:

Answering general questions

For answering general questions we often default to Perplexity. This is based on web search and can compile answers to questions based on the search results. It is fast, but will only give superficial answers to scientific questions.

TipGoogle search vs. LLM-assisted search

We most often use conventional search engines such as Google if we know what we search for (e.g. we want to look up opening hours of something). LLM-assisted search tools (Perplexity, but also ChatGPT with web search) are best in cases where we do not even exactly know what we search for.

Answering scientific questions

FutureHouse Platform

For surveying scientific papers, we often use the FutureHouse Plaform. On there, they have different agents.

Asta

An alternative tool, Asta, built by Ai2 can perform similar search and we often use it in combination with the FutureHouse Plaform.

OpenAI Deep Research (đź’µ)

Especially with GPT-5 Pro, Deep Research can sometimes deliver impressive results.

“Consuming papers”

For more easily digesting papers, we use different tools:

  • NotebookLM is excellent for converting paper(s) into forms that are easier to consume, such as podcasts, mindmaps, reports, videos. We mostly use it to generate podcasts and study guides. The podcasts are particularly useful for getting key ideas (or even differences) between papers in a form that can be consumed on the go. Sometimes the results are also useful to inspire a narrative.

  • Speechify can be useful to listen to papers (instead of reading them).

Guidelines