The Browser for Research.
LitLit is an AI research agent that turns papers into interactive knowledge maps. Discover key research, connect ideas, and explore what’s next.
Recursive self-improvement: AI systems that improve their own code, training or reasoningRecursive self-improvement has evolved from speculative foundations in the mid-20th century to concrete, empirically validated systems.
Early theory. I.J. Good’s 1965 paper 131 introduced the “intelligence explosion”: an ultraintelligent machine could design even better machines. Schmidhuber’s Goedel Machine 132 (2003) was the first rigorous framework: a system that rewrites its own code once it can prove the rewrite pays off. Nivel et al. 5 added bounds to keep it from degenerating.
The LLM era. STOP 4 showed an LLM improving its own “improver” code. Promptbreeder 148 and Gödel Agent 11 let agents mutate their own prompts and reasoning strategies.
Today. AlphaEvolve 129 scales the loop with evolutionary search, and Meta-Harness 142 has agents rewrite their own scaffolding to beat human-designed baselines.
Used by researchers at
See the Field, Not Just a List of Papers.
Search engines show you results. Citation maps show you links. LitLit helps you understand the research landscape—and navigate it.
- DINOv2: Learning Robust Visual Features without SupervisionCited by 11,212
- Scalable Diffusion Models with TransformersCited by 7,655
- RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic ControlCited by 4,698
- Diffusion policy: Visuomotor policy learning via action diffusionCited by 4,698
- OpenVLA: An Open-Source Vision-Language-Action ModelCited by 3,910
- Mastering Atari, Go, chess and shogi by planning with a learned modelCited by 2,791
World models for embodied AI: learning to predict and plan in latent spaceStart with a Question.
LitLit builds a map of the research around it. Ask follow-up questions, explore new directions, and watch the map grow.
Search beyond the words you know.
LitLit searches like a researcher: several queries, every candidate read and judged against your criteria, and the classics everyone cites pulled in on top.
- Understanding your question
- Searching the literature
- Reading and judging papers
- Finding foundational papers
- Placing papers on the map
- Finding research regions
- Naming regions
Find the papers. Understand the gaps.
Regions, lineages, core papers and rising work at a glance, plus the gaps between directions. Step out of the filter bubble and see the parts of the field you weren’t looking at.
Go Deeper, on the Map.
Ask anything. Answers cite the papers on the map, read full texts and figures when needed, and grow new branches that connect one direction to another.
Experimental setup. The experiments use superconducting transmon qubits in 2D grids with nearest-neighbour coupling, the native topology for the surface code.
Figure 1 · Quantum error correction below… — Surface code performance. shows the layout on Google’s Willow processor. Three code distances were tested on the same chip: distance-3, distance-5 and a single 101-qubit distance-7 code 1.
Follow an Idea Through Time.
Switch to Lineage and the same papers line up by year and influence. See where an idea started, who built on it, and which recent work is taking off.
What are you working on?
From exploring a new field to checking a research idea, start with the question on your mind.
Get oriented in a new field.
See the main research directions, the papers that shaped them, and where to start reading.
- Organizes the literature into clear research directions
- Highlights foundational papers, surveys, and emerging work
- Suggests a reading path from the basics to recent advances
World models for embodied AI: learning to predict and plan in latent spaceI recommend a reading path that moves from foundational concepts to modern architectures and finally to specialized applications.
- 145 Dream to Control (Hafner et al., 2019) — introduced learning latent dynamics for planning.
- 18 Learning Latent Dynamics for Planning from Pixels — lays the groundwork for latent dynamic models.
- 147 Mastering Diverse Domains through World Models — scales the approach up across domains.
- 149 V-JEPA 2 — a shift towards joint-embedding predictive architectures.
- 1Dream to Control: Learning Behaviors by Latent Imagination2019 · arXiv
- 2Learning Latent Dynamics for Planning from Pixels2018 · arXiv
- 3Mastering Diverse Domains through World Models2023 · arXiv
- 4V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning2025 · arXiv
- 5A Comprehensive Survey on World Models for Embodied AI2025 · arXiv
Follow the Evidence.
Cited, paper by paper
Every claim links to a paper on the map. Click a citation and the paper lights up.
Read, not just retrieved
Each paper is checked against your criteria before it lands on the map, so you can see why it is there.
Facts from the data
Authors, venues, years and citation counts come from the paper records, never from the model’s memory.
Built on open scholarly data
320M+scholarly works
2.9Bcitation links
180M+DOI records
Unpaywall50,000+sources of free copies
2.5M+preprints, read in full
See Research from a Different Angle.
Real maps made with LitLit. Pick one and start exploring
World models for embodied AI
Alzheimer’s disease beyond amyloid
AI for personalized cancer vaccines
Recursive self-improvement
Quantum error correction
Machine learning for new materials
Start free.
Every account gets free credits each month. Upgrade when you work through a field.
Free
Try LitLit on a question or two each month.
Pro
For working through a field: dozens of maps a month.
Max
For labs and heavy use, with room for deep maps every day.
Questions
How is this different from a search engine or a citation map?
A search engine gives you a ranked list for the keywords you typed. A citation map shows who cites whom around a paper you already have. LitLit is an agent: it runs the searches for you, reads and screens every paper against your question, organises the field by ideas, and then answers questions on the map, growing it when it runs out.
Which fields does it cover, and how fresh is the data?
Any field indexed by OpenAlex and Semantic Scholar: more than 300 million works across the sciences, medicine, engineering and the humanities, updated continuously. Preprints from arXiv are included.
Can I trust the answers?
Answers cite papers on the map, and paper facts come from the records, not from the model. LitLit can still misread a paper, so the citations are there for you to check. When a full text isn’t open access, it says the answer rests on the abstract.
Can it do a systematic review?
It helps with the searching, screening and organising: criteria-based screening of every paper, filters over the whole map, a saved library and exports to BibTeX, RIS and CSV. Protocol registration and reporting (e.g. PRISMA) stay with you.
Does it work with my reference manager?
Export BibTeX, RIS or CSV for Zotero, EndNote, Mendeley or anything else. Direct Zotero sync is coming soon.
What does LitLit remember, and who can see my maps?
Your maps are private unless you share a link. With memory on, LitLit remembers what you tell it about yourself, such as your field and how you like answers; you can see, edit or delete each memory, or turn memory off.
Break Out of the Search Box.
You don’t always know what to search for. That’s where exploration begins.