IAAILooped Actor
Looped Actor: Depth-Recurrent Reasoning Models for Reinforcement Learning
Brings dynamic looping to reinforcement learning: a transformer policy refines a latent toward a fixed point with a shared block and loops more in states that need planning. A complexity argument shows state-adaptive policies can need less expected computation than any fixed-runtime optimal policy.
Inside the method
State → [ Shared block ] × r(state) → fixed point → action
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- State-dependent loop count
- KV / state strategy
- Not applicable (policy network)
Reading note
Across 22 tasks in six environments, online and offline, it matches or beats an untied baseline with 16× more parameters. On Boxoban, loop counts grow with the remaining number of pushes.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopedactor2026,
title = {Looped Actor: Depth-Recurrent Reasoning Models for Reinforcement Learning},
author = {T. Konstantin Rusch and Tim Seyde and Jared Boyer and Zach J. Patterson and Daniela Rus},
year = {2026},
eprint = {2609.37432},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.37432}
}