In this post, I would like to offer evidence to support the following beliefs: 1. LLM search tools like Perplexity.ai greatly speed up the search portion of thinking about research questions. 2. These tools are more effective the more domain expertise the user has. 3. (As a consequence of 2.) New researchers should focus more on building expertise, not familiarity with AI tools.
I am a PhD student studying algebraic geometry. Over the last 6 weeks, I have been using Perplexity as a search engine to help me with my PhD research. I have been very impressed with how useful it is, and would like to share how I use it, and how it benefits me. Before continuing, I should disclose that everyone in my university did receive a 1-year free trial of Perplexity, which I have been using. I also want to mention that there are many similar tools, and I don’t think what I’m saying here is specific to Perplexity. See this post comparing AI research assistants.
Here I’d like to use Baron’s search-inference framework to think about thinking [1]. The search-inference framework splits the process of thinking up into searching for goals, possibilities, and evidence, and using the evidence to infer the probabilities that each possibility achieves each goal.
I find Perplexity greatly speeds up the search for evidence, and isn’t very helpful with inference. I haven’t had much opportunity to use it to search for possibilities, but I suspect it can do well there too.
Searching for Evidence with Perplexity
For me, thinking about mathematics research roughly looks like:
Find a goal, which can be a vague research goal or a specific lemma to prove.
Search for possible approaches that could solve the goal.
For each approach, look for evidence that it could succeed or fail. For example, this could be papers that use the approach to solve similar problems.
Sit down and try applying the most promising approach to solve the problem (I.e. infer if the approach solves the goal)
Repeat steps 2-4 until the goal is solved or I give up.
Perplexity has mostly helped me with step 3 so far.
An Example
Now let me give a specific example, and show how Perplexity helped me with step 3. The Grauert Direct Image Theorem tells us when there is an isomorphism between fibres of a flat family of coherent sheaves. In my current project, I am using this theorem to show two vector spaces are isomorphic, but the definition of the isomorphism is very abstract. I want to understand how the isomorphism is constructed, so I can apply it to a specific example.
After reading the relevant section in Hartshorne (Chapter III Section 12), I was still not clear on how to geometrically interpret this isomorphism. Thus, I used Perplexity to search for other people’s interpretations online. You can see the entire search session for yourself by clicking on this link.
The key part of the session came when I asked Can you find how people geometrically interpret the residue field at a point? To which I got the answer:
The residue field at a point in algebraic geometry has several geometric interpretations:
Natural codomain for functions: The residue field can be viewed as the natural codomain for functions defined at a given point on the scheme[3]. This interpretation aligns with the idea that the residue field represents the possible values that functions can take at the point.
Evaluation of sections: For any local section s ∈ 𝒪X(U) and any point x ∈ U, we can consider the “value of s at x” as sx = s̄ ∈ κ(x), where κ(x) is the residue field. This allows us to evaluate sections at points of the scheme[3].
Coordinates of points: The residue field is sometimes described as the “natural domain” for the coordinates of the point[2]. This interpretation is particularly relevant for closed points of varieties over algebraically closed fields.
Field of rational functions: For the generic point of a variety, the residue field can be interpreted as the field of rational functions on the variety[3]. This field consists of “coherent collections of values at closed points” that extend to compatible germs, which then glue to rational functions.
Fiber over a point: When considering morphisms between schemes, the residue field plays a crucial role in understanding the fiber over a point. For a morphism f: X → Y and a point y ∈ Y, the fiber f^(-1)(y) is closely related to the residue field κ(y)[5].
Geometric points: A morphism Spec(K) → X, where K is a field, is equivalent to giving a point x ∈ X and an extension K/κ(x)[2]. This interpretation allows us to think of geometric points as field-valued points of the scheme.
These interpretations collectively demonstrate how the residue field bridges the gap between the algebraic structure of schemes and their geometric properties, providing a crucial tool for understanding the local behavior of schemes at individual points[5].
The second interpretation stood out, as the morphism defined by evaluation of sections plays an important role in the problem I am studying. Before jumping into the references to understand exactly what was meant, I decided to ask for a bit more clarification: Point 2 in your response is interesting. If I have a coherent sheaf over a scheme, can I evaluate sections of that sheaf using residue fields?
In response to this, I got a mathematical explanation of how to use the residue field to evaluate sections. This is where it is important to be responsible – LLMs are not built to accurately do technical mathematics. However, in this case, I’ve seen this construction before, and I know enough to be comfortable that the answer is correct enough to accept. Critically when I wrote up this section of my paper, I went to the original sources and worked through the steps myself, to ensure that they’re correct.
Effective Use Requires Expertise
In my experience, this tool is more of a force-multiplier then something which solves problems on its own. The domain-expertise of the human operator is a key component of using Perplexity effectively. I find it most effective for two specific types of query, both very common in my research:
1. Reminding me of technical details.
Often I come across a situation where I know that I’ve learned some mathematical results which may be useful for my goal, but I can’t remember the specific details. Perplexity can very quickly find the exact statement in the online literature, much faster than I could look it up in a textbook or lecture notes myself.
2. Finding standard results.
Based on my mathematical intuition, I sometimes expect there to be a result of a specific form in the standard literature. In this case, I’ll ask Perplexity Are there any results for object X that relate property A and property B?, and if there is a standard result, it finds it quickly. Again, this is much faster than the alternative of looking in textbooks, online course notes and asking colleagues.
These two types of query essentially perfectly play into the strengths of LLMs, being that they read much faster than me, and are good at providing summaries of natural-language documents. These are also queries with very low risk of causing you to make mistakes in your work. This is because the LLMs essentially acts as a search engine, pointing you to the sources that you can then evaluate as you would if you found them on Google or in a library. Perplexity in particular seems to have some guardrails against hallucinations – in the search session about Grauert’s theorem above, I asked How does $phi_p$ (the isomorphism) act on a single element?, and got a very wordy reply that essentially said I don't know, I can't find it in the search results.
However both of the query types above require you to have some expertise in the subject area already! Without the years of math training I have, I would have no intuition and no previous knowledge suggesting what I should be asking the LLM.
Limitations & Room for Improvement
One caveat to all the praise I’ve given LLM search this far, is that it only really works for very well-known mathematics. If the result you need is “in the standard literature” then the LLM will find it. However, Perplexity’s search does not have access to the primary sources. The major reference textbooks, and more importantly, preprints and peer-reviewed publications, are not readily available to LLM search.
Instead, the results that Perplexity sorts through are human-made secondary material like course notes, presentation slides and StackOverflow posts. For more niche questions, the result you seek may be in the literature, but if nobody has posted about it online, it seems like Perplexity won’t find it.
In principle, the concept of LLM search could be combined with advances in machine-learning for topic modeling to build a really powerful tool for searching the literature. I am envisioning using LLM vectorizers and then tools like UMAP and HDBSCAN to get some semantic-embedding of all the papers on arXiv, and then using this as a database for the retrieval that Perplexity performs. Honestly, I would be surprised if this idea does not already exist, so please let me know if you’re aware of a tool like this!
Conclusion
If you do research and have some expertise in your domain, I believe you should give Perplexity (or a similar tool) a try to speed up the search portion of your research. Use it like you would Google or StackOverflow – be skeptical and verify what you find. Even if you’re LLM-skeptic, I think LLM powered search is one of the non-hype viable use cases for LLMs.
I also now more strongly believe that If you’re an aspiring researcher, you should focus on building your expertise. The tools will keep evolving, and who knows what the best tools will be when you’re a researcher yourself. Knowledge of tools is easy to obtain and gets obsoleted quickly. On the other hand, domain expertise and independent thinking skills are hard to obtain and probably won’t be obsoleted any time soon.
If you have different experiences with these tools, or think there is a way I could use them better, I’d love to hear what you think. Especially if you’ve had success using them on the parts of thinking that I believe they are less useful for; searching for goals, and evaluating the utility of various possibilities towards those goals.
Using LLM Search to Augment (Mathematics) Research
In this post, I would like to offer evidence to support the following beliefs:
1. LLM search tools like Perplexity.ai greatly speed up the search portion of thinking about research questions.
2. These tools are more effective the more domain expertise the user has.
3. (As a consequence of 2.) New researchers should focus more on building expertise, not familiarity with AI tools.
I am a PhD student studying algebraic geometry. Over the last 6 weeks, I have been using Perplexity as a search engine to help me with my PhD research. I have been very impressed with how useful it is, and would like to share how I use it, and how it benefits me. Before continuing, I should disclose that everyone in my university did receive a 1-year free trial of Perplexity, which I have been using. I also want to mention that there are many similar tools, and I don’t think what I’m saying here is specific to Perplexity. See this post comparing AI research assistants.
Here I’d like to use Baron’s search-inference framework to think about thinking [1]. The search-inference framework splits the process of thinking up into searching for goals, possibilities, and evidence, and using the evidence to infer the probabilities that each possibility achieves each goal.
I find Perplexity greatly speeds up the search for evidence, and isn’t very helpful with inference. I haven’t had much opportunity to use it to search for possibilities, but I suspect it can do well there too.
Searching for Evidence with Perplexity
For me, thinking about mathematics research roughly looks like:
Find a goal, which can be a vague research goal or a specific lemma to prove.
Search for possible approaches that could solve the goal.
For each approach, look for evidence that it could succeed or fail. For example, this could be papers that use the approach to solve similar problems.
Sit down and try applying the most promising approach to solve the problem (I.e. infer if the approach solves the goal)
Repeat steps 2-4 until the goal is solved or I give up.
Perplexity has mostly helped me with step 3 so far.
An Example
Now let me give a specific example, and show how Perplexity helped me with step 3. The Grauert Direct Image Theorem tells us when there is an isomorphism between fibres of a flat family of coherent sheaves. In my current project, I am using this theorem to show two vector spaces are isomorphic, but the definition of the isomorphism is very abstract. I want to understand how the isomorphism is constructed, so I can apply it to a specific example.
After reading the relevant section in Hartshorne (Chapter III Section 12), I was still not clear on how to geometrically interpret this isomorphism. Thus, I used Perplexity to search for other people’s interpretations online. You can see the entire search session for yourself by clicking on this link.
The key part of the session came when I asked
Can you find how people geometrically interpret the residue field at a point?To which I got the answer:The second interpretation stood out, as the morphism defined by evaluation of sections plays an important role in the problem I am studying. Before jumping into the references to understand exactly what was meant, I decided to ask for a bit more clarification:
Point 2 in your response is interesting. If I have a coherent sheaf over a scheme, can I evaluate sections of that sheaf using residue fields?In response to this, I got a mathematical explanation of how to use the residue field to evaluate sections. This is where it is important to be responsible – LLMs are not built to accurately do technical mathematics. However, in this case, I’ve seen this construction before, and I know enough to be comfortable that the answer is correct enough to accept. Critically when I wrote up this section of my paper, I went to the original sources and worked through the steps myself, to ensure that they’re correct.
Effective Use Requires Expertise
In my experience, this tool is more of a force-multiplier then something which solves problems on its own. The domain-expertise of the human operator is a key component of using Perplexity effectively. I find it most effective for two specific types of query, both very common in my research:
1. Reminding me of technical details.
Often I come across a situation where I know that I’ve learned some mathematical results which may be useful for my goal, but I can’t remember the specific details. Perplexity can very quickly find the exact statement in the online literature, much faster than I could look it up in a textbook or lecture notes myself.
2. Finding standard results.
Based on my mathematical intuition, I sometimes expect there to be a result of a specific form in the standard literature. In this case, I’ll ask Perplexity
Are there any results for object X that relate property A and property B?, and if there is a standard result, it finds it quickly. Again, this is much faster than the alternative of looking in textbooks, online course notes and asking colleagues.These two types of query essentially perfectly play into the strengths of LLMs, being that they read much faster than me, and are good at providing summaries of natural-language documents. These are also queries with very low risk of causing you to make mistakes in your work. This is because the LLMs essentially acts as a search engine, pointing you to the sources that you can then evaluate as you would if you found them on Google or in a library. Perplexity in particular seems to have some guardrails against hallucinations – in the search session about Grauert’s theorem above, I asked
How does $phi_p$ (the isomorphism) act on a single element?, and got a very wordy reply that essentially saidI don't know, I can't find it in the search results.However both of the query types above require you to have some expertise in the subject area already! Without the years of math training I have, I would have no intuition and no previous knowledge suggesting what I should be asking the LLM.
Limitations & Room for Improvement
One caveat to all the praise I’ve given LLM search this far, is that it only really works for very well-known mathematics. If the result you need is “in the standard literature” then the LLM will find it. However, Perplexity’s search does not have access to the primary sources. The major reference textbooks, and more importantly, preprints and peer-reviewed publications, are not readily available to LLM search.
Instead, the results that Perplexity sorts through are human-made secondary material like course notes, presentation slides and StackOverflow posts. For more niche questions, the result you seek may be in the literature, but if nobody has posted about it online, it seems like Perplexity won’t find it.
In principle, the concept of LLM search could be combined with advances in machine-learning for topic modeling to build a really powerful tool for searching the literature. I am envisioning using LLM vectorizers and then tools like UMAP and HDBSCAN to get some semantic-embedding of all the papers on arXiv, and then using this as a database for the retrieval that Perplexity performs. Honestly, I would be surprised if this idea does not already exist, so please let me know if you’re aware of a tool like this!
Conclusion
If you do research and have some expertise in your domain, I believe you should give Perplexity (or a similar tool) a try to speed up the search portion of your research. Use it like you would Google or StackOverflow – be skeptical and verify what you find. Even if you’re LLM-skeptic, I think LLM powered search is one of the non-hype viable use cases for LLMs.
I also now more strongly believe that If you’re an aspiring researcher, you should focus on building your expertise. The tools will keep evolving, and who knows what the best tools will be when you’re a researcher yourself. Knowledge of tools is easy to obtain and gets obsoleted quickly. On the other hand, domain expertise and independent thinking skills are hard to obtain and probably won’t be obsoleted any time soon.
If you have different experiences with these tools, or think there is a way I could use them better, I’d love to hear what you think. Especially if you’ve had success using them on the parts of thinking that I believe they are less useful for; searching for goals, and evaluating the utility of various possibilities towards those goals.
See Thinking and Deciding, Jonathan Baron, 2008