Launch HN: Undermind (YC S24) – AI agent for discovering scientific papers

292 points by jramette ↗ HN
Hey HN! We’re Josh and Tom from Undermind (https://www.undermind.ai/). We’re building a search engine for complex scientific research. There's a demo video at https://www.loom.com/share/10067c49e4424b949a4b8c9fd8f3b12c?..., as well as example search results on our homepage.

We’re both physicists, and one of our biggest frustrations during grad school was finding research — There were a lot of times when we had to sit down to scope out new ideas for a project and quickly become a deep expert, or we had to find solutions to really complex technical problems, but the only way to do that was manually dig through papers on Google Scholar for hours. It was very tedious, to the point where we would often just skip the careful research and hope for the best. Sometimes you’d get burned a few months later because someone already solved the problem you thought was novel and important, or you’d waste your time inventing/building a solution for something when one already existed.

The problem was there’s just no easy way to figure out what others have done in research, and load it into your brain. It’s one of the biggest bottlenecks for doing truly good, important research.

We wanted to fix that. LLMs clearly help, but are mostly limited to general knowledge. Instead, we needed something that would pull in research papers, and give you exactly what you need to know, even for very complex ideas and topics. We realized the way to do this is to mimic the research strategies we already know work, because we do them ourselves, and so we built an agent-like LLM pipeline to carefully search in a way that mimics human research strategies.

Our search system works a bit differently from casual search engines. First, we have you chat back and forth with an LLM to make sure we actually understand your really complex research goals up front, like you’re talking to a colleague. Then the system carefully searches for you for ~3 minutes. At a high level, it does something similar to tree search, following citation rabbit holes and adapting based on what it discovers to look for more content over multiple iterations (the same way you would if you decided to spend a few hours). The 3 minute delay is annoying, but we’re optimizing for quality of results rather than latency right now. At the end there’s a report.

We’re trying to achieve two things with this careful, systematic agent-like discovery process:

1. We want to be very accurate, and only recommend very specific results if you ask for a specific topic. To do this, we carefully read and evaluate content from papers with the highest quality LLMs (we’re just reading abstracts and citations for now, because they’re more widely accessible - but also working on adding full texts).

2. We want to find everything relevant to your search, because in research it’s crucial to know if something exists or not. The key to being exhaustive is the adaptive algorithms we’ve developed (following citations, changing strategy based on what we find, etc). However, one cool feature of the automated pipeline is we can track the discovery process as the search proceeds. Early on, we find many good results, and later on they get more sparse, until all the good leads are exhausted and we stop finding anything helpful. We can statistically model that process, and figure out when we’ve found everything (it actually has an interesting exponential saturation behavior, which you can read a bit more about in our whitepaper (https://www.undermind.ai/static/Undermind_whitepaper.pdf), which we wrote for a previous prototype.)

You can try searching yourself here:

133 comments

[ 4.4 ms ] story [ 234 ms ] thread
I have only tried one search, but so far it's impressive. I have been using elicit.com, but they seem to be taking a different approach that is less AI-heavy. I would definitely give this a shot for a few months.
We're trying to bias the system toward more autonomous execution, rather than a "copilot"-like experience where you iterate back and forth with the system. That lets us run more useful subroutines in parallel in the backend, as long as you specified your complex goal clearly.
This is really cool, excited to see where this goes!
OK, I'm both impressed and disappointed.

I did 2 searches.

First I asked about a very specific niche thing. I gave me results but none I wanted. It looked like I missed a crucial piece of information.

So I did the second search. I started with the final request it written for the previous search and added the information I though I missed. It gave me virtually the same results with a little sprinkle of what I was actually after.

A few observations:

1. I'm not sure but it seems like it relies too much on citation count. Or maybe citations in papers make it think that the paper is absolutely a must read. I specifically said I'm not interested in what's in that paper and I still got those results.

2. I don't see much dissertations/theses in the result. I know for sure that there a good results for my request in a few dissertations. None of them are in the results.

That said, while I didn't get exactly what I want I've found a few interesting papers even if they're tangential to the actual request.

A few possibilities: - We only use abstracts for now. Have to make sure you ask for something present there. - Did you ask for a scientific topic? (Sometimes people ask for papers by a specific author, journal, etc. The system isn't engineered to efficiently find that).

Regarding citations: we use them, but only for figuring out which papers to look at next in the iterative discovery process, not for choosing what to rank higher or lower at the end (unless you explicitly ask for citations). It's ranking based on topic match.

If you're comfortable, posting the report URLs here can let us debug.

This is a nice search rngine. I found it to be more effective than crawling with google scholar. Good work guys!
I’ve been using a similar platform that I really like called Answer This[1]. I’ll have to check out yours as well and see how it compares.

1. https://answerthis.io/

I ran these two example searches we have on our homepage on AnswerThis: (3D ion shuttling) https://undermind.ai/query_app/display_one_search/b3767fb7b6... (laser cooling to BEC) https://undermind.ai/query_app/display_one_search/c5f77f862a...

The results from their website aren't sharable, but their lists of references do not seem relevant (ie. they miss the fact that shuttling needs to be in 3D, and the list of experiments for laser cooling to BEC is missing all of the relevant papers).

I think, like other research tools, they're more focused on the summarization/extraction of information, rather than the discovery process (though they are similar to us in the way they say they do multi-stage retrieval and it takes some time).

AnswerThis founder here. Hang on tight, our next few updates will be instrumental in conversational feedback-based paper searches.

I love all the work that's being done in the field.

Can't wait for science to be faster.

Independent researcher without academic address; can't get in. Best of luck.
Very cool, and very relevant to my life -- I am currently writing a meta-analysis and finishing my literature search.

I gave it a version of my question, it asked me reasonable follow-ups, and we refined the search to:

> I want to find randomized controlled trials published by December 2023, investigating interventions to reduce consumption of meat and animal products with control groups receiving no treatment, measuring direct consumption (self-reported outcomes are acceptable), with at least 25 subjects in treatment and control groups (or at least 10 clusters for cluster-assigned studies), and with outcomes measured at least one day after treatment begins.

I just got the results back: https://www.undermind.ai/query_app/display_one_search/e5d964....

It certainly didn't find everything in my dataset, but:

* the first result is in the dataset.

* The second one is a study I excluded for something buried deep in the text.

* The third is in our dataset.

* The fourth is excluded for something the machine should have caught (32 subjects in total), but perhaps I needed to clarify 25 subjects in treatment and control each.

* The fifth result is a protocol for the study in result 3, so a more sophisticated search would have identified that these were related.

* The sixth study was entirely new to me, and though it didn't qualify because of the way the control group received some aspect of treatment, it's still something that my existing search processes missed, so right away I see real value.

So, overall, I am impressed, and I can easily imagine my lab paying for this. It would have to advance substantially before it was my only search method for a meta-analysis -- it seems to have missed a lot of the gray literature, particularly those studies published on animal advocacy websites -- but that's a much higher bar than I need for it to be part of my research toolkit.

For a meta-analysis, you might want to try the "extend" feature. It sends the agent to gather more papers (we only analyze 100 carefully initially), so if your report might say "only 55% discovered", could be useful.

(Also, if you want, you can share your report URL here, others will be able to take a look.)

(comment deleted)
For a systematic review/meta analysis you’d be expected to document your search strategy, exclusion criteria, etc anyway wouldn’t you? That’d preclude using a tool like this other than as a sense check to see if you needed to add more keywords/expand your search criteria anyway.

My wife does that for her day job (in the U.K. national healthcare system) and the systematic reviews have to be super well documented and even pre-registered on a system called PROSPERO. The published papers always have the full search strategy at the end.

I was planning to say "I used an AI search tool" and cite undermind.ai in my methods section. I think that won't raise any eyebrows in the review process but we'll see.
Have a look at the PRISMA reporting guidelines
I only have some experience writing normal papers, so just out of interest, could you elaborate what your usual search routine for a meta-analysis is?
There's a whole established process for this, see here for a textbook chapter https://training.cochrane.org/handbook/current/chapter-04

However, because I'm writing a methods-focused review -- we only look at RCTs meeting certain (pretty minimal) criteria relating to statistical power and measurement validity -- what I'm doing is closer to a combination of review of previous reviews (there have been dozens in my field) and a snowball search (searching bibliographies of papers that are relevant). I also consulted with experts in the field. however, finding bachelor's theses has been challenging, but many are actually relevant, so undermind was helpful there.

I'm a CS academic who _should_ be working on finalizing a new submission, so when I saw this on HN I decided to give it a try and see if it could find anything in the literature that I'd missed. Somewhat to my surprise, it did - the top 10 results contained two items that I really ought to have found myself (they're from my own community!), but that I'd missed. There were also some irrelevant results mixed in (and lots of things I was already aware of), but overall I'm very impressed with this and will try it out again in the future. Nice work :)
Any idea how i can use your tool for a vs code extension
Pretty good, it found some useful references I missed in Google Scholar and Arxiv. Looks promising, will use it more.
I've been using https://exa.ai for this. It doesn't do any advanced agent stuff like here, but it's way better than Google, especially if you're not quite sure what you're looking for.
Agreed, exa is great - particularly, it's the best thing I've found for fast web retrieval of slightly more complex topics than Perplexity, Google, etc can handle.
Are you planning to offer a search API at some point?
Potentially. Given the latency and the cost/compute we put into each result, it doesn't fit the usual API mechanics.

What use case are you thinking of?

How would you compare your product to elicit.ai?

In my opinion elicit has better looking UI and much more features and further along

I think the biggest difference is our focus on search quality, and being willing to spend a lot on compute to do it, while they focus on systematic extraction of data from existing sources and on being fast. It's a bit of an oversimplification (they of course have search, and we also have extraction).

Feature-wise, we definitely have a lot of work to do :) What crucial pieces do you think we're missing?

From what I understand, that’s not the case. They are working on both. I’d be concerned about how you can differentiate and compete with them. They have a big head start
In my experience, elicit’s big weakness is accuracy.
How is this different than the work that semantic scholar is doing around AI?
Semantic Scholar seems more focused on 1. being the data provider/aggregator for the research community, and 2. long term, I think they plan to develop software at the reading interface that learns as a researcher uses it to browse papers (a rich PDF reader, with hyperlinks, TLDRs, citation contexts, and a way to track your interactions over time, and remind you of what you've seen or not).

Their core feature now is a fast keyword search engine, but they also have a few advanced search features through their API (https://api.semanticscholar.org/api-docs/) like recommendations from positive/negative examples, but neither KW search nor these other systems are currently high enough quality to be very useful for us.

FYI our core dataset for now is provided by Semantic Scholar, so hugely thankful for their data aggregation pipeline and open access/API.

Do you plan on adding an API? I already have an inhouse knowledge discovery, annotation and search system that could be augmented by your service. Not super critical at this point, but a would be nice.

And yes, Semantic Scholar is a wonderful part of the academic commons. Fingers crossed they don't go down the jstor/oclc path.

I've used undermind for literature search and it was very precise! Thanks for the product! I wonder how you plan to extend the search to full paper content (will Semantic Scholar api allow this) - and do you plan to connect more datasets (which ones)? (many of them are paid...)
We'll certainly be able to include open access full texts, which is already a substantial fraction of the published papers, and a growing fraction too, as the publishing industry is rapidly moving toward open access. Paywalled full text search would require working with the publishers, which is more involved.
Great! I can definitely ask undermind for an overview paper of the scientific information landscape, unless you have a favourite in quick access to share?
Awesome! I just took you up on your offer and compared roughly similar questions using Claude 3.5 Sonnet and Undermind.

Claude 3.5 is reluctant to provide references—-although it will if coaxed by prompting.

Undermind solves this particular problem. A great complement for my research question —- the evidence that brain volume is reduced as a function of age in healthy cognitively normal humans. In mice we see a steady slow increase that averages out to a gain of 5% between the human equivalents of 20 to 65 years of age. This increase is almost linear as a function of log of age.

Here is the question that was refined with Undermind’s help:

>I want to find studies on adult humans (ages 20-100+) that have used true longitudinal repeated measures designs to study variations in brain volume over several years, focusing on individuals who are relatively healthy and cognitively functional.

I received 100 ranked and lightly annotated set of 100 citations in this format:

>[1] Characterization of Brain Volume Changes in Aging Individuals With Normal Cognition Using Serial Magnetic Resonance Imaging S. Fujita, ..., and O. Abe JAMA Network Open 2023 - 21 citations - Show abstract - Cite - PDF 99.6% topic match Provides longitudinal data on brain volume changes in aging individuals with normal cognition. Analyzes annual MRI data from 653 adults over 10 years to observe brain volume trajectories. Excludes populations with neurodegenerative diseases; employs true longitudinal design with robust MRI techniques.

It's worth highlighting that first result is exactly what you asked for, given all 4 of your criteria:

1. It's on adults.

2. It's longitudinal over multiple years.

3. It studies variations in brain volume.

4. It focuses on healthy individuals.

You can see the full results for that search text here: https://undermind.ai/query_app/display_one_search/e1a3805d35...

And that first hit in JAMA Open is a fabulous paper. Ten or mire yearly MRI scans for 650 subjects.
How is this different from Scite, Elicit, Consensus, and Scopus AI for Generating Literature Reviews
Ours is slow, but accurate, even for complex topics. The rest are fast, but generally can't handle complex topics. (There's more nuanced explanations in other comments)
I'll write the obligatory comment about doing literature searches in the 90s, which involved trudging to the physics library, the chemistry library, and the engineering library in search of dead tree copies of the journal articles you're after. Also: skimming each paper quickly after you photocopy it, to see if it references any other papers you should grab while you're at the library.
(comment deleted)
what are your biggest drawbacks?
Latency, compute required, and lack of full texts (paywalled publisher content).
Curious as to what it's doing under the hood, the query to return the results takes an excruciatingly long time... are you searching remote sources vs a local index?

this was the search <https://www.undermind.ai/query_app/display_one_search/cba773...> if you need a reference too it, ie bugs or performance monitoring...

The few minute time delay is primarily because of the sequential LLM processing steps by high quality LLMs, not database access times. The system reads and generates paragraphs about papers, then compares them, and we have to use the highest quality LLMs, so token generation times are perceptible. We repeat many times for accuracy. We find it's impossible to be accurate without GPT-4 level models and the delay.
would this be able to find the latest articles on a given topic?

let’s say i am interested in coffee and i’d like to get new research papers on it. would this work?

In short, yes, though it's geared toward topic search.

From a strategy perspective, we designed it for topic search because it makes more sense to find everything on a topic first, then filter for the most recent, if recent is what you want. That's because there is a lot of useful information in older articles (citation connections, what people discuss, and how), and gathering all that helps uncover the most relevant results. Conversely if you only ever filtered on articles in the last year, you might discover a few things, but you wouldn't have as much information to adapt to help the search work better.

So, you can ask for articles on coffee (though ideally it should be something a bit more specific, or there will be thousands of results). Our system will carefully find all articles, then you can filter for 2024 articles or look at the timeline.

Here is an open source tool for summarizing Arxiv papers: https://summarizepaper.com/
Very happy subscriber here, thank you for the tool. I do a lot of searching with it, however due to some changes in my life in the near future I will not need it as much so I wont be willing to spend $20 a month on it. So my question is, would you consider adding an option where one could pay per query rather than just per monthly subscription? I would love to use it for the occasional spark of curiosity when I want to know more about a certain topic without having to familiarise myself with the academic field surrounding it. Having a way for using undermind for situations like that would be truly amazing! Would gladly pay 1-2 or maybe even 3 dollars per extended query.
We've thought quite a bit about usage-based pricing, but found that doesn't work psychologically for most people. Generally, people seem happier by paying up front for access, then feel good about having the system available whenever they need it, rather then having to think through cost tradeoffs every time they want to do a search or use up credits. Please do reach out at support@undermind.ai though, we'd love to talk about a solution for you and get your feedback.
Have you considered the middle ground where there's a $10 tier ratelimited to X requests per timeperiod?
Not a bad idea. Want to avoid too much complexity in pricing, though. Decision fatigue.
This is really cool! Both of my parents are cell biologists, and I've done some time in labs as well, so a lot of paper exploring and reading in the family. "Review" articles are a good index but something more on-demand makes a lot of sense, I can definitely see this being extremely useful.