Trends over time

How open-endedness research shifts year over year, derived from 100 curated entries (from 2017 on): which concepts are rising, where citations concentrate, which models, datasets and tasks dominate, and how venues and categories change. Charts are interactive — hover for values, and click legend entries to toggle series.

Most-cited papers

The most-cited entries in the list (Semantic Scholar citation counts).

1 Voyager: An Open-Ended Embodied Agent with Large Language Models · Preprint (arXiv) 2023 1,830
2 The AI Scientist: Towards fully automated open-ended scientific discovery · Preprint (arXiv) 2024 865
3 Eureka: Human-Level Reward Design via Coding Large Language Models · ICLR 2023 630
4 Genie: Generative Interactive Environments · Preprint (arXiv) 2024 629
5 MineDojo: Building Open-Ended Embodied Agents with Internet-Scale Knowledge · NeurIPS 2022 579
6 Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution · Preprint (arXiv) 2023 490
7 Emergent Complexity and Zero-shot Transfer via Unsupervised Environment Design · NeurIPS 2020 315
8 Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions · GECCO 2019 299
9 The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search · Preprint (arXiv) 2025 269
10 Debating with More Persuasive LLMs Leads to More Truthful Answers · Preprint (arXiv) 2024 269
11 Evolutionary Optimization of Model Merging Recipes · Nat Mach Intell (2025) 2024 234
12 Open-Ended Learning Leads to Generally Capable Agents · Preprint (arXiv) 2021 229
13 Prioritized Level Replay · ICML 2021 217
14 Automated design of agentic systems · ICLR 2025 215
15 Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Prompts · NeurIPS 2024 198

What the field says comes next

The forward-looking sections of every paper — “future work”, “conclusion”, “limitations”, “open questions” — mined for the curated concepts authors flag as open. Ranked by how many papers propose each direction; ▲ recent marks concepts where most of those calls land in the last few years (rising momentum).

Open-ended learning 17
Large language models (LLMs) ▲ recent 13
Exploration 8
Unsupervised environment design 7
POET / paired open-ended 6
Reinforcement learning ▲ recent 5
Quality-diversity ▲ recent 5
MAP-Elites 4
Automated scientific discovery ▲ recent 4
Emergent abilities 3
Recursive self-improvement ▲ recent 3
Meta-learning 3

Charts over time

Share of each year's entries engaging each curated open-endedness concept (matched across full text) — watch the shift toward LLM / agent work.

Share of each year's papers whose “future work / conclusion / limitations” sections explicitly propose each curated concept (detected only in those forward-looking sections). Concepts that papers keep naming as open — increasingly so — are the emerging hot topics.

Share of each publication-year's total citations captured by each concept — weights every entry by how often it's cited, so it tracks where impact concentrates, not merely how many papers appear.

Share of each year's entries mentioning a model / algorithm family (keyword-based, approximate) — the LLM families climb fast.

Share of each year's entries mentioning an AI lab / model provider (keyword-based, approximate).

Share of each year's entries mentioning a dataset / benchmark / environment (keyword-based, approximate).

Share of each year's entries mentioning a task type (keyword-based, approximate).

Primary arXiv categories of the papers each year — which subfields drive open-endedness (from arXiv metadata, not keywords).

Share of each year's papers that link to code — a reproducibility signal. Blog posts and videos are excluded since they can't ship a repository.

Median (bars) and mean (line) citations per paper by publication year. Recent years are necessarily under-counted — citations accrue with age.

How the leading themes wax and wane year over year.

Linear trend of each consensus keyphrase's yearly mentions — green is rising, red is fading.

How the leading venues' yearly counts stack up (the arXiv-preprint share shows how much work stays unpublished).

Entries by curated category each year.