Interoceptive Attention as Dynamic Homeostatic Prioritization in a Foraging Agent
Published in From Animals to Animats 18 (Springer), pp. 192–203. Main-track talk in Berlin on Wednesday 21 October 2026 at 15:50, with a lightning talk and poster at the WWIN workshop on Monday 19 October.
Abstract
Biological systems must regulate competing needs under limited perceptual bandwidth, where sharpening one estimate costs the capacity to sharpen the others. Any fixed-budget system therefore has to decide where to allocate its perceptual precision. We study this in a foraging agent that must keep several bodily needs satisfied to survive, modelled with active inference. At each step it reads its own body-state beliefs, identifies the most-needed channel, and reallocates a fixed budget of interoceptive precision toward it, so that the same precision-shaped likelihood feeds both belief update and planning. In AffectWorld, a four-channel foraging gridworld, this selective allocation more than doubles learning-phase survival at matched budget against a uniform-precision agent (0.414 vs 0.199 across 11 layouts, n = 32 seeds each, paired cluster-bootstrap p ≤ 10⁻⁴). Two further results sharpen the mechanism. The benefit runs through planning as well as perception, since denying the shaped likelihood to the planner alone removes about half of it. It is also need-aligned, since aiming precision at the least-needed channel does worse than spreading it evenly. The attended channel additionally learns its own dynamics about twice as fast, and stays ahead even at matched observation count, a behavioural trace of the same precision routing, visible in learning speed, not survival.
Notes on the paper
A few details that are easy to miss or that I would now state differently. These are my own notes, not an official erratum. Section A of the arXiv supplement gives the full as-run settings.
- The planner scored policies on the pragmatic term only, with no ambiguity or state-information-gain term. κ therefore reached planning through the predicted observations, not through both EFE terms as Sec. 2.1 says.
- All runs used policy precision γ = 1 (the library default), not γ = 16 as stated in Sec. 2.1.
- The ground-truth selector mentioned in Sec. 3.2 is the single-layout pilot (n = 3) from Sec. 2.2 and is not shown in Fig. 4.
- At α₀ = 10 the inference-only ablation falls to the level of the uniform agent (0.023 against 0.000), not the anti-aligned agent (0.492) as Sec. 3.3 says. The 88-point drop stands.
- The learning-speed result (“about twice as fast”) is a difference in level that is present from the first trial, rather than faster convergence.