Tim Rocktäschel
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,656 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,656 words
Add missing citations, update stale details, or suggest a clearer explanation.
Tim Rocktäschel is a German computer scientist known for his work on reinforcement learning, open-ended learning, and language-based AI agents. He is a Professor of Artificial Intelligence at University College London (UCL), where he leads the Deciding, Acting, and Reasoning with Knowledge (DARK) Lab, and until 2026 he was a Director and Principal Scientist at Google DeepMind, where he led the Open-Endedness team [1][2]. In 2026 he co-founded Recursive Superintelligence, a London- and San Francisco-based startup that aims to build self-improving AI [3][4]. His research output is widely cited; he is a co-author of the original retrieval-augmented generation (RAG) paper and a co-creator of the NetHack Learning Environment, and he shared two Best Paper Awards at the 2024 International Conference on Machine Learning [1][5][6].
| Item | Detail |
|---|---|
| Born | Berlin, Germany |
| Field | Artificial intelligence, reinforcement learning, open-endedness [1] |
| Doctorate | PhD, Computer Science, University College London (2017) [1] |
| Doctoral advisor | Sebastian Riedel [1] |
| Current academic role | Professor of Artificial Intelligence, UCL (since 2023) [1][2] |
| Notable industry role | Director / Principal Scientist and Open-Endedness Team Lead, Google DeepMind (2022 to 2026) [1] |
| 2026 role | Co-founder, Recursive Superintelligence [3][4] |
| Best-known work | RAG, NetHack Learning Environment, Genie [5][6][7] |
Rocktäschel grew up in Berlin, where he completed his Abitur at the Heinrich-Hertz Oberschule in 2006 with advanced courses in mathematics and computer science [1]. He studied at the Humboldt-Universität zu Berlin from 2006 to 2012, earning a Diplom in Informatik (computer science), the German degree roughly equivalent to a master's, with a subsidiary field in psychology [1]. His diploma thesis on jointly extracting proteins and biomolecular events using factor graphs won a faculty award, and during this period he placed first in the SemEval 2013 DDIExtraction shared task, an early sign of his interest in extracting structured knowledge from text [1].
He moved to UCL for doctoral study, completing his PhD in computer science between 2013 and 2017 [1]. His thesis, titled "Combining Representation Learning with Logic for Language Processing," was supervised by Sebastian Riedel, with Daniel Tarlow and Thore Graepel as additional advisors [1]. The doctorate was supported by a Microsoft Research PhD Scholarship and, later, a Google PhD Fellowship in natural language processing [1].
After his PhD, Rocktäschel spent 2017 and 2018 at the University of Oxford. He was a postdoctoral researcher in reinforcement learning, a Junior Research Fellow at Jesus College, and a Stipendiary Lecturer at Hertford College, where he tutored undergraduates in functional and imperative programming [1].
In August 2018 he joined Facebook AI Research (FAIR), the lab now part of Meta AI, in London, while simultaneously taking up a faculty post at UCL [1]. He rose from research scientist to lead the lab's reinforcement learning team, serving as a Research Manager and Area Lead before leaving in 2022 [1]. His time at FAIR coincided with several of his most-cited works, including the RAG paper and the NetHack Learning Environment, both published in 2020 [5][6].
Rocktäschel joined Google DeepMind in May 2022 as a Senior Staff Research Scientist and lead of the Open-Endedness team [1]. He had previously interned at DeepMind in 2015 during his doctoral studies [1]. In November 2024 he was promoted to Director and Principal Scientist while continuing to lead the team [1].
The Open-Endedness team studied systems that can keep generating new tasks and improving indefinitely, rather than converging on a fixed objective. Rocktäschel has framed open-endedness as a route toward artificial general intelligence and self-improvement, organized around three threads: training agents that set their own goals, building large-scale world models that supply agents with endless environments, and connecting these methods to large language models so that AI can improve itself [2]. During this period he was a co-author of Genie, an 11-billion-parameter foundation world model that learns to generate playable, action-controllable environments from unlabeled internet videos [7].
Rocktäschel has held a UCL appointment continuously since 2018, when he started as an Assistant Professor (Lecturer) at the Centre for Artificial Intelligence in the Department of Computer Science [1]. He was promoted to Associate Professor in 2021 and to full Professor of Artificial Intelligence in 2023, giving his inaugural lecture in 2024 [1][8]. He is the principal investigator of the DARK Lab, whose name stands for Deciding, Acting, and Reasoning with Knowledge, and his group's work on agents, world models, and self-improvement mirrors the themes of his DeepMind research [2].
He has been recognized within the broader European research community as a Scholar and later a Fellow of the European Laboratory for Learning and Intelligent Systems (ELLIS) [1]. He has also served the field as a senior area chair at ICML, NeurIPS, and ICLR, and helped organize a series of workshops on agent learning in open-endedness and on language in reinforcement learning [1].
In 2026 Rocktäschel co-founded Recursive Superintelligence, a startup pursuing recursively self-improving AI [3][4]. The company is led by Richard Socher, the former chief scientist at Salesforce and founder of the search engine You.com, as chief executive, with Rocktäschel as a co-founder drawing on his DeepMind background in open-endedness and world models [3][9]. Reporting placed its founding team, which also includes researchers who previously worked at OpenAI, Meta, and Uber AI, as having come together over late 2025 and early 2026 [9].
In May 2026 the company emerged from stealth, having raised more than 650 million dollars at a valuation of about 4.65 billion dollars [3][4]. The round was led by GV (formerly Google Ventures) and Greycroft, with participation from the chipmakers Nvidia and AMD through their venture arms [3][4]. Recursive is incorporated in London with offices in London and San Francisco and a team of fewer than thirty people [4]. Its stated plan is to first build AI that improves AI and then extend the approach to other scientific fields, automating the cycle of generating, testing, and validating research ideas [3].
Rocktäschel's Google Scholar profile lists more than 42,000 citations and an h-index above 60, with a research focus tagged simply as open-endedness [5]. His work falls into a few broad areas.
His early work combined neural networks with logic and structured knowledge. "Reasoning about Entailment with Neural Attention" (ICLR 2016), written with Edward Grefenstette, Karl Moritz Hermann, Tomáš Kočiský, and Phil Blunsom, introduced a word-by-word neural attention mechanism for natural language inference and was an early, influential demonstration of attention for textual entailment [10]. "End-to-End Differentiable Proving" (NeurIPS 2017), an oral presentation, showed how to perform differentiable theorem proving by combining the structure of symbolic provers with learned vector representations of symbols [5]. He also contributed to "Language Models as Knowledge Bases?" (EMNLP 2019), which probed how much factual knowledge is stored in the parameters of pretrained language models [5].
Rocktäschel is the tenth-listed author of "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (NeurIPS 2020), the FAIR paper led by Patrick Lewis that introduced RAG [11]. RAG combines a pretrained sequence-to-sequence generator with a learned retriever over an external document index, fine-tuned end to end, and the approach became a standard technique for grounding language models in retrieved evidence [11]. It is by a wide margin his most-cited paper [5].
He co-created the NetHack Learning Environment (NeurIPS 2020), a fast, procedurally generated reinforcement learning testbed built on the roguelike game NetHack [6]. The authors argued that NetHack is complex enough to drive long-term research on exploration, planning, and skill acquisition while remaining cheap to simulate, making it a useful benchmark for hard-exploration and open-ended learning [6]. He later helped author a related framework, MiniHack, for designing custom RL tasks on the same engine [1]. His survey work in this area includes "A Survey of Zero-Shot Generalisation in Deep Reinforcement Learning" (Journal of Artificial Intelligence Research, 2023), a widely cited review of how RL agents generalize to unseen situations [5].
At DeepMind his research turned to generative world models and LLM-driven self-improvement. He shared two ICML 2024 Best Paper Awards: one for "Genie: Generative Interactive Environments," the foundation world model that generates controllable environments from video, and one for "Debating with More Persuasive LLMs Leads to More Truthful Answers," a study of using debate between language models to elicit truthful answers [7][12]. He was also a co-author of "Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution" (ICML 2024), which evolves and improves the prompts that guide a language model [5].
Beyond the two ICML 2024 Best Paper Awards, Rocktäschel has received a Best Paper Award at the Conference on Automated Knowledge Base Construction in 2020 for "How Context Affects Language Models' Factual Predictions," a Best Paper Award at the EMNLP SocialNLP workshop in 2016 for "emoji2vec," and outstanding and highlighted area chair awards at ICLR in 2021 and 2022 [1]. His doctoral funding came from a Microsoft Research PhD Scholarship and a Google PhD Fellowship, and he was elected a Junior Research Fellow at Jesus College, Oxford [1].