IAAIFlexLoop
FlexLoop: Depth-Elastic Looped Policies for Adaptive Test-Time Computation in Deep RL
Finds that pretrained looped RL policies specialize to their full trained depth. FlexLoop post-trains them with the original RL objective plus adjacent-depth policy distillation, making shallower depths reliable and enabling state-wise adaptive inference.
Inside the method
[ Shared block ] × R → distill depth d+1 into depth d
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Depth-elastic, state-wise adaptive
- KV / state strategy
- Not applicable (policy network)
Reading note
On 30 long-horizon goal-conditioned environments it keeps full-depth performance, cuts average recurrent depth by up to 43% and gives up to 1.34× wall-clock speedup in a stress test.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{flexloop2026,
title = {FlexLoop: Depth-Elastic Looped Policies for Adaptive Test-Time Computation in Deep RL},
author = {Xun Wang and Ruishuo Chen and Yu Chen and Zhuoran Li and Longbo Huang},
year = {2026},
eprint = {2609.34488},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.34488}
}