We've upgraded! Visit our new site at www.undermind.ai. The new deep search tool covers all of science, not just ArXiv, and is 10x faster. Your previous data is still available at app.undermind.ai.

Deep Search with an AI Agent

Powered by GPT-4, Undermind searches all of arXiv, covering physics, math, ML, and more

Undermind outperforms keyword search by 10-50x

The Problem

Finding scientific information is hard, requiring lots of manual filtering.

The Solution

Our end-to-end system carefully collects precisely the relevant works in reponse to a complex search topic.

The Team

Undermind was started by two MIT quantum physics PhDs, Thomas Hartke and Joshua Ramette.

Team photo

How Undermind Finds Research

Imagine asking a colleague to find papers on a scientific topic for you

The Undermind search agent:

  1. Understands your complex search topic.
  2. Searches and finds precisely relevant papers on arXiv, reading deep within full texts, and adapting based on what it discovers.
  3. Explains exactly why each paper found matches your search topic, so you can understand and trust its decisions.
  4. Statistically estimates the total number of works done on your topic, to make sure you've found everything.

Undermind is much more than:

  • A keyword search engine. Instead, Undermind reasons about the content of papers to decide relevance, and is insensitive to phrasing.
  • A retrieval augmented generation (RAG) chatbot. Instead, Undermind adaptively searches as part of a complex, many-step pipeline.
  • Abstract search. Instead, we use full texts and semantic embeddings from the 2.3 million articles on arXiv for maximum search quality.

Example Search Reports

You can generate reports on almost anything, but here we highlight a few key use cases

Understand exactly what others have done on a topic

Gather a comprehensive list of papers in a specific field, including key results and concepts, to develop an expert's perspective:
  1. Tokenization-free large language model architectures, which have been shown to achieve compute/accuracy tradeoffs comparable to or better than traditional token-based models
  2. Efficient, high-fidelity approximations of dense feedforward layers in transformers, such as low-rank approximations and parameterizations incorporating structured sparsity
  3. Experimental demonstrations of microwave to optical conversion for interconnecting superconducting qubits
Test novelty to see if a topic has been highly explored (routing trapped ions in 2D) or is still fairly novel (routing trapped ions in 3D), to understand the impact of potential projects, or determine whether it's possible to patent an idea or method:
  1. Experimental realizations of routing or shuttling trapped ions in three dimensions in a quantum computer
  2. Experimental realizations of routing or shuttling trapped ions in two dimensions in a quantum computer

Find the resources that help you learn about new topics

Look for pedagogical introductions and overviews of topics. Our system will filter for papers that provide introductory and clear explanations, rather than advanced explanations:
  1. Introductions to Lieb-Robinson bounds in quantum systems
Find articles with clear explanations of detailed and specific concepts. Our system will highlight articles that actually explain topics, rather than just mention them in passing:
  1. Why are bang-bang protocols in quantum control time optimal?
Find all review articles on a topic:
  1. Reviews of rydberg atom quantum computing

Find solutions to problems you're working on

Find explanations of specific methods:
  1. Systematic quantitative analyses of large language model performance optimization in a server inference setting, with a focus on how to navigate latency/throughput tradeoffs when processing large numbers of incoming requests
  2. Strategies for weighting different noise levels when training diffusion models. In particular, any work proposing strategies for determining either (1) how often to sample each noise level during training, or (2) how heavily to weight the loss associated with each noise level when computing the overall loss
  3. Explanations of how to implement modulation transfer spectroscopy, including details about the error signal and electronic circuits
  4. Methods to multiplex control of semiconductor quantum dot spin qubits in dense architectures
Explore open-ended solutions to problems:
  1. How to cool the spin excitations in a Fermi-Hubbard gas without using evaporative cooling

Dive into any particular niche subtopic

Understanding Search Reports

The database: the ArXiv preprint repository

  • We have all research published on the online repository ArXiv since its inception in 1991. This is around 2.3 million papers, updated daily whenever new papers are posted. Thank you to arXiv for use of its open access interoperability.
  • Thank you also to Semantic Scholar for providing citation data, and the Semantic Scholar Open Research Corpus (S2ORC), a preprocessed database of many of the full texts from many arXiv PDFs.

Understanding how Undermind comprehensively gathers all papers on a topic

Our agent systematically explores the literature like a human scientist, following citation trails, and adapting where it looks based on what it has found so far. By tracking this process, we can construct a statistical model, and thereby determine when all the papers have likely been found.

The more specific your topic, the faster the system converges to find everything. It generally follows an exponential form:
Search Sizes Visualization
Partway through the search process, we can predict the likely outcome of a search (how many total relevant papers would be found if we analyzed the entire database) using the statistical model:
Paper Discovery Visualization

Example of statistical model of an example search process. Left: Number of relevant papers we found vs. the number of papers the language model has read. Because we look at the most promising papers first, there are diminishing returns. Blue lines extending from black dot show samples of statistical predictions of how many papers might be discovered, if we spend more resources reading further papers. Red line shows number of relevant papers actually found with additional computation. Right: Probability distribution over the actual number of relevant papers that would be discovered if the database was exhaustively searched. As more papers are read, the distribution tightens.