Sebastian Risi

Research profile

Research summary

Sebastian Risi, a researcher at Sakana AI in Japan, works at the junction of evolutionary computation, artificial life, and deep-learning architecture design. The OpenAlex snapshot records an h-index of 34 and an i10-index of 90, with 49 works in the most recent three-year window; the countsByYear series peaked at 32 publications in 2020, dipped to 11 in 2023, and climbed to 21 in 2026. Topical coverage is unusually broad: 138 works tagged Evolutionary Algorithms and Applications, 98 Artificial Intelligence in Games, 68 Reinforcement Learning in Robotics, 57 Neural Networks and Applications, and 30 Modular Robots and Swarm Intelligence, with the two assigned fields split evenly (0.5/0.5) between computational biology (artificial life / morphogenesis modeling) and machine learning (deep learning architectures). Two 2026 papers reviewed by Evolve mark the current agenda: an LoRA-NCA study showing reusable regulatory handles control phenotypic variation — praised as elegant with strong simulation evidence, though universality and biological claims were judged to overreach — and a temporal-recurrence study arguing streaming models can match deep networks with fewer layers, with solid sweeps but language-model results lacking seeds and unproven throughput gains. Radar scores capture the shape: activity 100, focus 0, collaboration 41, reproducibility 55.

Research keywords

Highlights

• Reusable regulatory handles steer phenotypic variation

  Evolve's review of this 2026 LoRA-NCA study (middle author) calls it elegant: phenotypic scale maps onto low-rank regulatory directions in weight space, with strong simulation evidence — while flagging that universality and biological claims overreach. Reproducibility readiness is rated medium.

• Temporal recurrence as an alternative to depth

  In this 2026 paper (last author), the claim is that recurrence lets streaming models match deep networks with fewer layers, spanning recurrent Transformers, mixture of experts, and deep reinforcement learning. Evolve finds the sweeps solid but notes the language-model results lack seeds and throughput gains are unproven; readiness medium.

• Sustained, broadening output across five topic clusters

  Beyond the two reviewed papers, the snapshot shows 49 works in the last three years and a 2020 peak of 32 publications, spread across evolutionary algorithms (138), AI in games (98), RL in robotics (68), neural networks (57), and modular robots and swarm intelligence (30).

Timeline

1. 2008First recorded output appears: 1 publication in 2008, growing to 5 by 2010.
2. 2015Output climbs to 17 publications, opening a sustained high-productivity stretch.
3. 201620 publications, followed by 21 in both 2017 and 2018 — a stable plateau.
4. 2020Peak year: 32 publications recorded in a single year.
5. 2023Output dips to 11 publications before recovering.
6. 202621 works recorded so far, including both Evolve-reviewed papers: the LoRA-NCA morphogenesis study (middle author) and the temporal-recurrence architecture paper (last author).

Papers in the Evolve catalog

• 2026-09-24

  On Growth and Form, and Function: Reusable Regulatory Handles Control Phenotypic Variation

• 2026-09-11

  Temporal Recurrence Favors Fewer Layers

Frequent co-authors