RL-trained RNNs show hybrid fixed-point plus quasi-periodic attractor dynamics, whereas SL-trained RNNs on the same task converge almost exclusively to fixed-point-only solutions

measured in 1 paper

- Kononov, Pospelov, Anokhin, Nekorkin & Maslennikov train 250-unit vanilla RNNs from scratch on a context-dependent decision-making (CtxDM) task using either reinforcement learning (PPO) or supervised learning (Adam + cross-entropy) on the identical task, then analyze each trained network's dynamics by numerical fixed-point finding. [kononov-etal-2025-hybrid-computational-dynamics-rl] - RL-trained RNNs show hybrid dynamics: "stable fixed points, where neural activity converges to a single static pattern, and quasi-periodic attractors" (Results), whereas SL-trained RNNs on the same task converge almost exclusively to fixed-point-only landscapes. The quasi-periodic attractors are reported in the paper's terms; mapping them to a limit-cycle node is not asserted here. [kononov-etal-2025-hybrid-computational-dynamics-rl]

Structure

Fixed-point graph topology (population-level universality of RNN dynamical structure)

Context

numerical fixed-point reconstruction of trained RNN dynamics, RL vs. supervised-learning training-signal effect on attractor-landscape complexity

Confirmed in models

Kononov et al. 2025 CtxDM ReLU-RNN (250 units, RL- and SL-trained) · Kononov et al. 2025 Context-Dependent Decision-Making RNN

Method

Papers

Reinforcement learning gives rise to emergent hybrid computational dynamics in recurrent neural networks — Kononov, Roman A., Pospelov, Nikita A., Anokhin, Konstantin V., Nekorkin, Vladimir V., Maslennikov, Oleg V.2025 · arXiv:2510.11162