The Eleventh International Conference on Learning Representations (ICLR) ran from May 1 to May 5, 2023, in Kigali, Rwanda, the first major AI conference hosted in Africa. Google DeepMind announced on April 27, 2023, that its teams would present 23 papers there, grouped around general-purpose learning, new methods for reasoning and reinforcement learning, and AI for science. The meeting has since moved on to Vienna (2024), Singapore (2025), and Rio de Janeiro (2026), where Google reported more than 95 papers.
Key takeaways
- ICLR 2023 was held May 1 to May 5, 2023, in Kigali, with a hybrid format that allowed remote participation.
- DeepMind listed 23 papers and sponsored the event as a Diamond sponsor and DEI champion.
- Highlighted work ranged from a test of 2,200 models on formal-language tasks to a simulator for colliding objects.
- The figures come from DeepMind's own preview post, so they describe what the lab chose to feature, not an independent ranking.
- Three editions have followed: Vienna in May 2024, Singapore in April 2025, and Rio de Janeiro in April 2026, where Google listed more than 95 papers.
Why the 2023 venue mattered
ICLR is a yearly meeting on deep learning and representation learning, the part of machine learning that studies how models build useful internal descriptions of data. The 2023 edition took place at a convention venue about five kilometers (three miles) from Kigali International Airport, and combined an in-person program with remote viewing for people who could not travel. Conference content, including videos, was scheduled to become free to the public on June 3, 2023.
DeepMind's post describes the meeting as the first major AI conference hosted in Africa and the first in-person event of its kind since the pandemic began. The post says DeepMind supported the conference as a Diamond sponsor and DEI champion, which refers to sponsorship tiers and diversity, equity, and inclusion programs.
The 23 papers, by theme
The post does not reprint all 23 papers. It selects a handful and sorts them into three groups. The table restates the highlighted items in plain terms; every entry is as DeepMind described it, and none was independently tested for this article.
| Theme | Highlighted work | What DeepMind reported |
|---|---|---|
| Open questions on general learning | Generalization and the Chomsky hierarchy | 2,200 models were tested on 16 tasks; some struggled to generalize, and adding external memory was needed for improvement |
| Open questions on general learning | Learning to solve a problem from two perspectives | Training a model to view one problem in two ways improved generalization across similar tasks |
| Open questions on general learning | Long-horizon, sparse-reward tasks | A new approach and an open-source training dataset helped models explore over long time spans |
| New methods | Explainable multi-step reasoning | A method draws on the logical structure inside a language model to produce explanations people can check |
| New methods | Data-efficient reinforcement learning | A method matched the Agent57 benchmark level on Atari with 200 times less experience |
| AI for science | Molecular property prediction and quantum chemistry | A denoising approach set a new state of the art on molecular tasks; a transformer improved quantum chemistry calculations from atomic positions alone |
| AI for science | FIGnet simulator | A physics-inspired model of collisions between complex shapes, such as a teapot and a doughnut |
Reading the two most concrete results
The formal-language study is the easiest to interpret. The Chomsky hierarchy ranks computational problems by the kind of memory a machine needs to solve them. Testing 2,200 models on 16 tasks asks which model designs can handle which levels. The reported finding, that some models failed to generalize and that external memory was central to progress, speaks to a long-running question about what neural networks can learn from examples alone.
The reinforcement learning result is worth reading with its baseline in mind. Agent57 was an earlier DeepMind system that needed nearly 80 billion frames of game experience to reach human-level scores across 57 Atari games. A method that reaches the same performance with 200 times less experience points to cheaper training, although Atari games are a narrow test and the post does not claim the gain carries over to other domains.
The FIGnet item shows why simulation matters outside research. Models that predict how irregular objects collide can feed robotics, graphics, and mechanical design, the applications the post names.
What changed since: ICLR 2024 to 2026
The conference has kept its yearly rhythm and moved across three continents after Kigali. ICLR 2024, the twelfth edition, took place May 7 to May 11, 2024, in Vienna, Austria. ICLR 2025, the thirteenth, ran April 24 to April 28, 2025, at the Singapore EXPO. The fourteenth edition was held April 23 to April 27, 2026, at the Riocentro Convention and Event Center in Rio de Janeiro, Brazil.
| Edition | Dates | Venue |
|---|---|---|
| ICLR 2023 (11th) | May 1 to May 5, 2023 | Kigali, Rwanda |
| ICLR 2024 (12th) | May 7 to May 11, 2024 | Vienna, Austria |
| ICLR 2025 (13th) | April 24 to April 28, 2025 | Singapore |
| ICLR 2026 (14th) | April 23 to April 27, 2026 | Rio de Janeiro, Brazil |
For 2026, Google's event page says it was a Diamond sponsor, with researchers from Google Research and Google DeepMind contributing, and lists more than 95 papers, seven oral presentations, and 24 workshops. It features a booth talk, a kiosk demo, and an EXPO session from DeepMind's Robotics Vision team on 3D foundational robotics models. Google publishes a combined Google Research and DeepMind count, so it cannot be compared directly with the 23-paper DeepMind figure from 2023. No official ICLR 2026 submission or acceptance totals were found on the conference page.
Limits of this record
DeepMind's post is a conference preview, written before the meeting began. It does not report reviews, citation counts, or later adoption of any paper. The ICLR Outstanding Paper Awards for 2023 were announced on a separate ICLR page, and that page lists four winning papers without author affiliations, so no claim about DeepMind winning an award is made here. Readers who want the complete 23-paper list should use the schedule DeepMind links from its post, and anyone judging a specific result should read the paper itself.
FAQ
What does ICLR stand for?
International Conference on Learning Representations. The 2023 edition was the eleventh.
When and where was ICLR 2023?
May 1 to May 5, 2023, in Kigali, Rwanda, with remote participation available.
Where was ICLR held after 2023?
Vienna, Austria, in May 2024; Singapore in April 2025; and Rio de Janeiro, Brazil, from April 23 to April 27, 2026.
How many papers did DeepMind present?
DeepMind's April 27, 2023, post says its teams presented 23 papers.
Did DeepMind win an ICLR 2023 outstanding paper award?
The ICLR announcement lists four winning papers without author affiliations, so this could not be confirmed from the sources checked and is not claimed.
Related reading
- How AlphaGo learned to play Go
- DeepMind learning resources
- AI in scientific research
- Whether AI can replace scientists
More in DeepMind.
Last verified: October 2026
Sources
- DeepMind's latest research at ICLR 2023, Google DeepMind, April 27, 2023
- ICLR 2023 conference page, International Conference on Learning Representations, 2023
- Announcing the ICLR 2023 Outstanding Paper Award Recipients, ICLR blog, 2023
- ICLR 2024, ICLR 2025, and ICLR 2026 conference pages, International Conference on Learning Representations
- Google at ICLR 2026, Google Research, 2026