Work

The planning loop needs a controller: a neuromodulated, learned budget for simulation

Year: 2026
Type: preprint
Abstract: A prefrontal–hippocampal planning architecture (Jensen et al. 2026, eLife, “spacetime attractor”; Jensen et al. 2024) pairs a fast amortised planner in prefrontal cortex with an explicit hippocampal simulator, the simulator teaching the planner. This note argues the loop requires a component it does not specify: a controller setting, per trial, how many simulations run, how widely they sample, and whether a familiar solution is reopened for search. A gradient-based stopping rule cannot distinguish a converged plan from a stale model after an unobserved environmental change. The note proposes that the controller is neuromodulatory (noradrenaline raising budget and sampling temperature on unexpected uncertainty; dopamine gating consolidation), driven in part by the loop’s own prefrontal–hippocampal disagreement, and itself learned. Three testable predictions are given, with prior art including the 2025–2026 LC→OFC noradrenaline learning-rate results, from which the claim is distinguished. Part 1 is written for a general reader; Part 2 for specialists. Idea arose during the author’s attendance at K. T. Jensen’s OIST TSVP talk, 2026-11-19. Files: v0.1 text, its OpenTimestamps proof, and the OpenTimestamps proof of the frozen v0 (priority record).
Language: English
Open Access status: green
OpenAlex ID: W7236816691
Cites: 2
Cited by:
Related to: 10