MATH · IN · MODELS

Kononov et al. 2025 Context-Dependent Decision-Making RNN

Research (Kononov, Pospelov, Anokhin, Nekorkin & Maslennikov 2025; custom-trained from scratch, RL vs. SL, on the CtxDM cognitive task)

Structures found in this family (1)

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

By model (1)

Kononov et al. 2025 CtxDM ReLU-RNN (250 units, RL- and SL-trained) · 250 units
Fixed-point graph topology (population-level universality of RNN dynamical structure)

Observations (1)

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

Papers

Reinforcement learning gives rise to emergent hybrid computational dynamics in recurrent neural networks (2025)