Source-linked AI summary
Two Routes to Interoceptive Prioritisation: Dissociating Likelihood Precision and Preference Reweighting in Active Inference
St John Grimbly, Nicolas Kuske, Evert A. Boonstra, Charel van Hoof, Rowan Hodson, Benjamin Rosman, Ryan Smith, Mark Solms, Jonathan P. Shock
TL;DR
Competing bodily needs require an agent to prioritize while inferring their urgency, but behavior alone may not reveal the mechanism. The paper compares likelihood-precision and preference interventions in a foraging agent, separately and together. On additional layouts both improve average survival and their interaction is positive, while internal measurements show effects survival alone cannot resolve.
Problem
Agents must prioritize competing bodily needs, and reproducing behavior alone does not identify whether prioritization relies on interoceptive gain or preferred outcomes.
Source: §1 · p. 1, §5 · p. 10Method
The paper compares likelihood-precision routing and preference reweighting, separately and together, in an active-inference forager that learns its body model online.
Source: §3.2 · p. 4, §3.3 · p. 5Results
Across 72 additional layouts, both interventions improve average survival and their interaction is positive, but the predicted preference benefit on difficult layouts was not established.
Source: §5 · p. 10, §6 · p. 11Takeaways & Limitations
Uncertainty or gain, distance from physiological limits, and experienced state range can be measured alongside behavior when evaluating prioritization models.
Source: §5 · p. 10Preference reweighting does not target oxygen directly, so the comparison does not test symmetric access to every need channel.
Source: §3.3 · p. 5
Abstract
from arXiv · showhide
An embodied agent must prioritise among competing bodily needs. In active inference, it can respond to an urgent need by increasing the precision of information about that need or by increasing its preference for outcomes that relieve it. We compare these two interventions, separately and together, in a foraging gridworld with three lethal needs, and ask how each changes survival, perception and the body model the agent learns. Precision sharpens interoceptive posteriors and improves survival. Preference reweighting keeps bodily states further from their lethal boundaries and changes the states from which the agent learns, so its learned model covers fewer need levels, with lower quality across all levels but slightly higher quality on the levels it visits. An experiment on 72 further layouts does not establish the predicted preference benefit on difficult layouts, but both interventions improve average survival there and their interaction is positive. Survival alone therefore gives an incomplete account of how a model prioritises needs and learns about its body.
1 Introduction
Agents facing competing bodily needs must prioritize among them while inferring their urgency. This paper compares distinct precision- and preference-based interventions, asking how each affects behavior, learning, and biological interpretation.
1 Introduction
- A multi-need agent cannot service every drive at once and must infer urgency from interoceptive signals to prioritize survival.The paper frames active inference as modeling the relationship between bodily-state inference and homeostatic control.Source: §1 · pp. 1–2
- Likelihood-precision routing makes a selected need’s signal more reliable, whereas preference reweighting makes outcomes relieving that need more attractive.The first draws on precision-weighted sensory evidence; the second on motivational and homeostatic accounts.Source: §1 · p. 2
- The interventions are distinct: precision changes bodily observations and their likelihood, while preferences change sought outcomes and thereby the experience available for learning.Preferences do not directly enter the likelihood update, but favored actions shape what the agent experiences.Source: §1 · p. 2
- The same behavior can arise from different mechanisms, so reproducing an animal’s behavior does not identify whether prioritization uses interoceptive gain or preferred outcomes.Precision is linked to gain control on interoceptive afferents, while preference is linked to changes in regulated outcomes.Source: §1 · p. 2
- The study tests precision and preference interventions separately and together, relating survival to uncertainty, bodily regulation, and learned state coverage.It also tests whether preference effects depend on layout difficulty using an independently selected set of layouts.Source: §1 · p. 2
2 Related Work
Prior work models interoceptive attention, preferences, and related routes to behavioral change, while this paper varies likelihood precision and preference as distinct interventions.
2 Related Work
- Smith et al. distinguish interoceptive sensory precision from a bodily-state prior; this paper instead compares likelihood precision with preferences.A state prior expresses a belief about the body, whereas a preference specifies desired observations.Source: §2 · p. 2
- Precision-as-attention accounts describe attention through the precision of sensory evidence or prediction errors; a discrete-state visual-search model has used context-dependent likelihood precision.Source: §2 · p. 2
- Motivational and homeostatic accounts encode desirable bodily or contextual outcomes through prior preferences, while homeostatic reinforcement learning uses reduced physiological deficit as reward.Source: §2 · p. 3
- Other active-inference accounts derive valence from policy precision or couple preferences with attentional precision.The present comparison makes its two interventions separately controllable within a discrete active-inference framework.Source: §2 · p. 3
- Earlier work studied selective precision allocation [9], preference variants [26], and preference effects on likelihood learning, but not this paired interoceptive comparison.This paper compares how each intervention changes the experience from which the agent learns.Source: §2 · p. 3
3 Methods
The study places an active-inference forager with three lethal needs in a resource gridworld and contrasts likelihood-precision routing with preference reweighting. It measures survival, perception, bodily regulation, and learned body-model coverage across original and additional layouts.
3.1 A foraging task with three lethal needs in tension
- The 6×6 forager couples hunger, thirst, and oxygen to food and water resources, with drinking relieving thirst while lowering oxygen and death at any lethal boundary.Food and water tiles remain fixed and reusable.Source: §3.1 · pp. 3–4, Figure 1 · p. 3
3.2 An active-inference forager that learns its body model online
- A discrete POMDP agent infers three noisy interoceptive need levels, selects policies by expected free energy, and updates only its interoceptive likelihood online.No condition selects actions to improve its body model.Source: §3.2 · p. 4
3.3 Likelihood precision and preference reweighting
- κ-routing reallocates fixed likelihood precision across channels, changing emitted evidence and inference, whereas dynamic-C rewrites food and water preferences for planning without directly changing A.Precision allocation necessarily de-prioritises other channels when one channel receives more precision.Source: §3.3 · p. 5, Figure 2 · p. 6
- Because dynamic-C does not directly target passive oxygen while κ-routing can prioritize it, the comparison does not test symmetric access to all channels.Oxygen recovers on land and has no navigable resource cue.Source: §3.3 · p. 5
3.4 What changes in behaviour and learning?
- Evaluation separates survival, posterior entropy, learned-likelihood quality across all versus visited need levels, visited-state breadth, and true-state minimum headroom.Minimum headroom measures the most endangered need’s distance from its lethal boundary, scaled to that channel’s range.Source: §3.4 · p. 6
3.5 Experiments and analysis
- The original study uses 11 layouts and 8 paired seeds; a separately fixed 72-layout test targets preference’s difficult-layout survival effect, while original intervals are wide.Original runs comprise 20 learning episodes of 120 steps, with learned-model comparisons over a longer run.Source: §3.5 · p. 7
4 Results
Precision routing and the combined intervention improve survival on the original layouts, while preference reweighting changes bodily regulation and the states represented in the learned model. On additional layouts, the predicted preference benefit on difficult layouts was not established, although average survival effects and their interaction were positive.
4.1 How does each intervention change survival?
- Survival was 0.56 at baseline, 0.83 with κ-routing, 0.67 with preference reweighting, and 0.94 with both interventions on the original layouts.Precision and the combined intervention improved survival after multiple-comparison correction; preference alone remained uncertain, with an interval spanning decreases and increases.Source: §4.1 · pp. 7–8, Table 1 · p. 8
4.2 Precision sharpens perception, preference changes the learned model
- κ-routing sharpened interoceptive posteriors; preference reweighting visited fewer need levels, lowering all-level model quality despite slightly higher quality on visited levels.The preference contrast for posterior entropy was smaller and its interval included zero; precision showed no comparable drop in all-level quality.Source: §4.2 · pp. 8–9, Table 3 · p. 9
4.1 How does each intervention change survival?
- On 72 additional layouts, the predicted preference benefit on difficult layouts was not established, though both interventions had positive average survival effects and a positive interaction.On these layouts, κ-routing lowered posterior entropy, while dynamic-C raised headroom and lowered conflict.Source: §4.1 · p. 8, Table 2 · p. 9
4.3 Preference keeps needs further from their lethal boundaries
- Preference raised minimum headroom by +0.05 (+0.11 combined); both interventions cut simultaneous near-lethal conflict from about 2% of baseline steps to less than 0.1%.Precision sharpened evidence about need, whereas preference kept the most endangered need further from its boundary.Source: §4.3 · p. 9, Table 3 · p. 9
4.4 Robustness on the original layouts
- Preference's easy-layout survival pattern persisted with true body levels, alternative penalty settings, and held-out seeds, but did not generalize to additional layouts.Replacing posterior deficits with true body levels left preference survival unchanged on every original layout.Source: §4.4 · p. 9
- On six original layouts, belief-only κ-routing improved survival by +0.33, but this was not a paired decomposition of the full intervention's +0.27 contrast across eleven layouts.The subset result suggests sharper emission is not required for a benefit, while the primary intervention remains composite.Source: §4.4 · pp. 9–10
5 Discussion
Likelihood precision and preference reweighting affect different parts of prioritisation, perception and learning, so survival alone cannot distinguish their mechanisms. On additional layouts both interventions improved average survival, but the predicted preference benefit on difficult layouts was not established.
5 Discussion
- Precision lowers posterior entropy, while preference raises it; preference also raises headroom and narrows coverage, whereas precision slightly lowers headroom.On additional layouts, precision slightly lowered headroom; neither intervention is confined to perception or action.Source: §5 · p. 10
- Preference narrows the bodily states visited, leaving rarely visited levels close to the prior, while visited levels are learned no less accurately.An average over the whole likelihood can therefore obscure what the agent learned from the states it actually encountered.Source: §5 · p. 10
- Across the 72 additional layouts, both interventions improved average survival and their interaction was positive, but preference did not help on calibrated difficult layouts.No continuous relation with baseline survival was detected, and the unadjusted tertile contrasts ran in the opposite direction from the original-layout pattern.Source: §5 · p. 10
- Behavioural fit alone cannot identify whether prioritisation works through interoceptive gain or preferred outcomes; internal measurements can help distinguish them.The paper compares implemented mechanisms rather than formally testing parameter identifiability.Source: §5 · p. 10
- The original experiment used eleven selected layouts with a post-hoc difficulty split, and preference had no direct oxygen target because oxygen recovers passively.The joint hunger–thirst gate was never triggered; broader channels, richer environments and longer planning horizons would test how far the findings extend beyond this gridworld and three-step planner.Source: §5 · pp. 10–11
6 Conclusion
The paper compares likelihood precision and preference reweighting as distinct mechanisms for interoceptive prioritisation. The predicted preference benefit on independently calibrated difficult layouts was not established, while both interventions improved average survival across the 72 layouts and their interaction was positive.
6 Conclusion
- The predicted preference benefit on independently calibrated difficult layouts was not established.Across all 72 layouts, secondary analyses found positive average survival effects for both interventions and a positive interaction.Source: §6 · p. 11
- The two mechanisms have different dominant effects on posterior uncertainty, bodily regulation and state coverage.Their internal observables reveal effects on later inference and learning that survival alone does not resolve.Source: §6 · p. 11
- Comparing internal observables alongside behaviour is important when using behavioural success to evaluate a model of prioritisation.The paper concludes that survival alone does not resolve how the mechanisms combine or affect later inference and learning.Source: §6 · p. 11
A Model and constants
The model allocates interoceptive precision to the believed most urgent need and reweights preferences for food and water from believed deficits. On the original layouts, preference alone helped on lower-baseline-survival layouts and reduced survival on the others.
A Model and constants
- The model represents hunger and thirst with latent levels 0–10 and oxygen with levels 0–6, using six ordinal need observations per channel.The cited passage identifies the channel levels and observation count.Source: A Model and constants · p. 14
- Under a precision budget of K = 1.95, the attended channel receives κ = 0.9 and each other channel receives 0.525, versus 0.65 per channel under uniform allocation.The attended channel is the one believed closest to its lethal boundary, with ties broken by lowest index.Source: A Model and constants · p. 14
- Dynamic-C sets food and water preferences from believed hunger and thirst deficits, with reference level 15 and critical threshold τ = 14.A deficit exceeds the threshold only when the expected need level is below 1; oxygen is not reweighted, and C does not enter the likelihood update.Source: A Model and constants · pp. 14–15
- Table S1 gives constants that are identical across the four conditions.The caption identifies the table's scope but does not specify individual constant values.Source: Table S1 · p. 15
- Preference alone improved survival by +0.36 on five lower-baseline-survival layouts and reduced it by −0.11 on the other six.The overall survival contrast had a 90% interval of [−0.03, +0.24], which does not establish equivalence within ±0.06.Source: A Model and constants · p. 15
B Additional layouts
The additional-layout test did not find the predicted preference benefit on the hard tier, and the marginal preference effect increased across the unadjusted middle and easy tertiles. Internal measures also differed between precision routing and dynamic-C.
B Additional layouts
- The 72 layouts were selected without reference to intervention outcomes, calibrated into three groups of 24, and evaluated on eight new seeds.They were distinct under grid rotation and reflection and from the original layouts.Source: B Additional layouts · p. 16
- The marginal preference effect was −0.008 on hard layouts, +0.059 [+0.003, +0.122] on middle layouts and +0.190 [+0.128, +0.257] on easy layouts.The hard-tier mean was the primary test; its seed-median estimate was −0.011, with leave-one-layout-out means from −0.018 to −0.006.Source: B Additional layouts · p. 16, Figure S1 · p. 16
- Marginal κ-routing lowered posterior entropy, broadened coverage and slightly lowered headroom, while marginal dynamic-C raised entropy and headroom, lowered conflict and narrowed coverage.Each listed effect had a 95% interval excluding zero, except κ-routing's effect on conflict.Source: B Additional layouts · p. 16
C State coverage and the learned-model measure
Preference reweighting changes which need levels the agent visits, so fewer likelihood columns are updated; the levels it visits are learned at least as well. The reported quality comparison averages across levels and comes from a separate learning run.
C State coverage and the learned-model measure
- Preference reweighting leads the agent to visit fewer levels while learning the levels it does visit at least as well.Preference does not enter the Dirichlet update directly; changed visitation affects which likelihood columns are updated.Source: C State coverage and the learned-model… · p. 17
- The comparison uses a separate 30-episode, 150-step learning run, not the main text’s three-layout run with all four conditions.In the separate run, hunger and thirst ranged up to 15 rather than 10; the main-text comparison with precision used an earlier run of the same runner.Source: C State coverage and the learned-model… · p. 17
- Table S2 summarizes quality averaged over all levels of three channels across 96 paired layout–seed records on 12 layout IDs, with intervals resampling layout IDs.Two layout IDs share coordinates.Source: Table S2 · p. 17