arxiv: 2604.24527 · v1 · submitted 2026-04-27 · 💻 cs.AI

Recognition: unknown

Interoceptive machine framework: Toward interoception-inspired regulatory architectures in artificial intelligence

Diego Candia-Rivera (NERV)

Authors on Pith no claims yet

Pith reviewed 2026-05-08 03:40 UTC · model grok-4.3

classification 💻 cs.AI

keywords interoceptionembodied AIadaptive autonomyhomeostasisallostasisenactive cognitioninternal state regulationartificial agents

0 comments

The pith

The interoceptive machine framework abstracts biological internal-state regulation into three principles to design more adaptive AI architectures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes the interoceptive machine framework to translate principles of monitoring and regulating internal signals from biology into computational designs for artificial agents. It organizes these into homeostatic regulation of internal viability, allostatic anticipatory re-evaluation based on uncertainty, and enactive generation of data through interaction. The goal is to embed internal state variables and regulatory loops so that AI systems handle decision-making, uncertainty, and interactions more robustly in changing environments. This approach treats the principles as functional abstractions rather than literal biological copies, aiming to support greater autonomy in embodied AI. Potential uses include improved human-computer interaction and assistive technologies.

Core claim

The central claim is that abstracting interoception into homeostatic, allostatic, and enactive principles supplies a unifying perspective for embedding internal-state regulation and regulatory loops in AI architectures, which in turn supports more robust decision-making, calibrated uncertainty handling, and adaptive interaction strategies in uncertain and dynamic environments.

What carries the argument

The interoceptive machine framework, which maps the monitoring, integration, and regulation of internal signals to three functional principles that assign distinct computational roles: internal viability regulation, anticipatory uncertainty-based re-evaluation, and active data generation through interaction.

If this is right

AI systems gain robust decision-making through embedded internal state variables and regulatory loops.
Calibrated uncertainty handling emerges from the allostatic principle of anticipatory re-evaluation.
Adaptive interaction strategies improve via the enactive principle of active data generation.
Agents achieve functionally grounded self-regulation applicable to embodied AI in dynamic settings.
Direct implications arise for human-computer interaction and assistive technologies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The framework could be extended to supply internal reward signals in reinforcement learning by tying them to homeostatic viability metrics.
Implementation in simple robotic testbeds could reveal whether the three principles reduce sample inefficiency in long-horizon tasks.
The approach might connect to existing work on predictive processing by treating allostatic re-evaluation as a form of active inference.
If validated, the framework suggests a route for designing AI that maintains functional awareness of its own operational limits.

Load-bearing premise

Abstracting biological interoception into the three functional principles will produce effective computational architectures for AI without needing direct neurophysiological mappings or immediate empirical validation of the resulting agents.

What would settle it

Implementing the three principles in embodied AI agents and testing them against standard agents in uncertain dynamic environments would falsify the framework if the interoceptive agents show no measurable gains in decision robustness, uncertainty calibration, or adaptive interaction.

read the original abstract

This review proposes an integrative framework grounded on interoception and embodied AI-termed the interoceptive machine framework-that translates biologically inspired principles of internal-state regulation into computational architectures for adaptive autonomy. Interoception, conceived as the monitoring, integration, and regulation of internal signals, has proven relevant for understanding adaptive behavior in biological systems. The proposed framework organizes interoceptive contributions into three functional principles: homeostatic, allostatic, and enactive, each associated with distinct computational roles: internal viability regulation, anticipatory uncertainty-based re-evaluation, and active data generation through interaction. These principles are not intended as direct neurophysiological mappings, but as abstractions that inform the design of artificial agents with improved self-regulation and context-sensitive behavior. By embedding internal state variables and regulatory loops within these principles, AI systems can achieve more robust decision-making, calibrated uncertainty handling, and adaptive interaction strategies, particularly in uncertain and dynamic environments. This approach provides a concrete and testable pathway toward agents capable of functionally grounded self-regulation, with direct implications for human-computer interaction and assistive technologies. Ultimately, the interoceptive machine framework offers a unifying perspective on how internal-state regulation can enhance autonomy, adaptivity, and robustness in embodied AI systems

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A clear conceptual synthesis of interoception into three computational principles for AI design, but it remains untested at the level of abstractions.

read the letter

The main takeaway is that this paper organizes interoception into a tripartite framework—homeostatic for internal viability, allostatic for anticipatory re-evaluation, and enactive for active data generation through interaction—and maps each to roles in building self-regulating AI agents. It positions the whole thing as a testable design pathway rather than a finished system. That structure is the actual new element: a compact way to pull together scattered ideas from biology and embodied cognition into one set of guidelines for internal-state loops in artificial systems. The authors stay careful by calling these abstractions, not direct brain copies, which avoids the usual overreach in bio-inspired work. The writing keeps the focus on practical implications for decision-making under uncertainty and for things like human-computer interaction. The soft spot is straightforward: there are no implementations, no simulations, no code, and no comparisons to existing methods like predictive processing or homeostatic reinforcement learning. The claim that embedding these principles will produce more robust agents therefore rests on the unverified hope that the biological translation will be effective once someone tries to code it. That gap is real but not fatal for a review-style proposal; it just means the paper functions as an organizing lens rather than a ready-to-use result. This is worth bringing to a reading group for anyone working on embodied or adaptive AI who wants a fresh way to think about internal regulation. It will not change core methods on its own, but it could help frame design choices in uncertain environments. I would send it to peer review in a venue like Adaptive Behavior or Frontiers in Neurorobotics because the synthesis is coherent and the topic has clear downstream uses, even if referees will need to push for more concrete next steps.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes the interoceptive machine framework, an integrative abstraction that organizes biological interoception into three functional principles (homeostatic, allostatic, and enactive) and maps them to distinct computational roles—internal viability regulation, anticipatory uncertainty-based re-evaluation, and active data generation through interaction—for designing embodied AI systems with enhanced self-regulation, calibrated uncertainty, and adaptive behavior in dynamic environments. The work positions itself as inspirational rather than a direct neurophysiological mapping and frames the framework as a testable pathway with implications for human-computer interaction and assistive technologies.

Significance. If the proposed abstractions prove effective when instantiated, the framework could supply a unifying biologically grounded perspective for improving autonomy and robustness in embodied AI, particularly by embedding internal-state variables and regulatory loops. The manuscript is credited for its explicit avoidance of direct biological-to-computational mappings and for presenting the ideas as a forward-looking, testable direction rather than an empirically validated result.

major comments (1)

[Abstract] Abstract: The statement that the framework 'provides a concrete and testable pathway' is not supported by any specific architectural specifications, pseudocode, algorithms, or illustrative agent implementations in the manuscript, rendering the pathway more abstract than claimed and limiting its immediate actionability for AI design.

minor comments (2)

The manuscript would benefit from a short section or appendix containing at least one high-level pseudocode sketch or block diagram showing how, for example, the allostatic principle could be realized as an uncertainty-driven re-evaluation module in a standard reinforcement-learning agent.
The discussion of implications for human-computer interaction and assistive technologies remains high-level; adding one or two concrete scenarios (e.g., an interoceptive controller for a wearable health monitor) would strengthen the applied relevance without requiring new experiments.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive and positive review, which recognizes the framework's potential as a unifying perspective for embodied AI. We address the single major comment below and will incorporate revisions to improve clarity.

read point-by-point responses

Referee: [Abstract] Abstract: The statement that the framework 'provides a concrete and testable pathway' is not supported by any specific architectural specifications, pseudocode, algorithms, or illustrative agent implementations in the manuscript, rendering the pathway more abstract than claimed and limiting its immediate actionability for AI design.

Authors: We agree that the manuscript presents a high-level conceptual framework rather than a fully specified implementation. The phrase 'concrete and testable pathway' was intended to convey that the three abstracted principles (homeostatic, allostatic, and enactive) supply distinct, implementable guidelines for embedding internal-state regulation in AI agents, which can then be empirically evaluated in future work. However, we acknowledge that the wording may overstate immediacy in the absence of pseudocode or examples. We will revise the abstract to clarify that the framework 'outlines a conceptual pathway toward testable implementations' of regulatory architectures, better reflecting its inspirational and abstraction-focused scope. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper is a conceptual review that abstracts interoception from biological literature into three functional principles (homeostatic, allostatic, enactive) and proposes their mapping to computational roles in AI architectures. No equations, fitted parameters, predictions, or derivations are present that reduce to the paper's own inputs by construction. The framework is explicitly positioned as inspirational abstractions and a testable pathway rather than a validated or self-referential result. No self-citation chains, uniqueness theorems, or ansatzes are invoked in a load-bearing manner that would create circularity. The central claims remain independent of any internal fitting or renaming of known results.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The proposal rests on the domain assumption that interoception principles can be usefully abstracted for AI without loss of functional value, plus the background premise that internal-state regulation improves autonomy in uncertain environments.

axioms (1)

domain assumption Interoception, as monitoring, integration, and regulation of internal signals, is relevant for understanding adaptive behavior in biological systems.
Explicitly stated as the grounding premise in the opening of the abstract.

pith-pipeline@v0.9.0 · 5507 in / 1243 out tokens · 50408 ms · 2026-05-08T03:40:24.324711+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

10 extracted references · 4 canonical work pages

[1]

gut-feeling

The interoceptive machine framework within the AI theoretic landscape 5.1 Philosophical positions around enactivism and AI. The enactivist and machine-autonomy literatures encompass a range of positions. Strong autopoietic enactivism grounded in the biological theory ties autonomy to metabolic self-production [6,88], and therefore, more inclined to the si...
[2]

Evaluation of interoceptive architectures should rely on measurable system properties rather than anthropomorphic interpretations

Practical considerations 6.1 Evaluation paradigms and testable predictions The proposed architecture generates concrete empirical predictions that can be evaluated in AI systems. Evaluation of interoceptive architectures should rely on measurable system properties rather than anthropomorphic interpretations. Each of the proposed principles may produce dis...
[3]

Internal regulatory signals can become unintended optimization targets, creating misalignment between internal system objectives and externally defined assistance goals

Limitations assessment 7.1 Failure modes of interoception-inspired architectures Introducing interoceptive or self-monitoring variables into artificial agents may improve certain properties, but it also introduces distinctive risks that require careful consideration. Internal regulatory signals can become unintended optimization targets, creating misalign...
[4]

Autonomous systems must generate their own identity and determine the significance of their interactions

A path toward near-conscious AI? As developed in previous sections, functionally grounded, context-sensitive behavior, the process by which an agent organizes behavior relative to internal and external constraints, requires more than just embodied interaction with the environment. Autonomous systems must generate their own identity and determine the signi...
[5]

being by doing

Conclusion Embodied AI, despite its advances, still falls short of capturing the essential qualities of human-like intentionality and meaning. A fundamental ontological gap persists between artificial systems and living organisms. Most notably, artificial systems lack a self-sustaining, self-producing mode of existence, often referred to as "being by doin...
[6]

Acknowledgements The author thanks Marie-Constance Corsi for the valuable feedback to an early version of this work
[7]

O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S

References [1] Park JS, O’Brien J, Cai CJ, Morris MR, Liang P, Bernstein MS. Generative Agents: Interactive Simulacra of Human Behavior. Proc. 36th Annu. ACM Symp. User Interface Softw. Technol., New York, NY, USA: Association for Computing Machinery; 2023, p. 1–22. https://doi.org/10.1145/3586183.3606763. [2] Sap M, Le Bras R, Fried D, Choi Y. Neural The...

work page doi:10.1145/3586183.3606763 2023
[8]

Mechanistic versus phenomenal embodiment: Can robot embodiment lead to strong AI? Cogn Syst Res 2001;2:251–62

Sharkey NE, Ziemke T. Mechanistic versus phenomenal embodiment: Can robot embodiment lead to strong AI? Cogn Syst Res 2001;2:251–62. https://doi.org/10.1016/S1389-0417(01)00036-5. [17] Anderson ML. Embodied Cognition: A field guide. Artif Intell 2003;149:91–130. https://doi.org/10.1016/S0004-3702(03)00054-7. [18] Froese T, Ziemke T. Enactive artificial in...

work page doi:10.1016/s1389-0417(01)00036-5 2001
[9]

A survey on neuro-mimetic deep learning via predictive coding

Salvatori T, Mali A, Buckley CL, Lukasiewicz T, Rao RPN, Friston K, et al. A survey on neuro-mimetic deep learning via predictive coding. Neural Netw 2026;195:108161. https://doi.org/10.1016/j.neunet.2025.108161. [70] Kendall A, Gal Y. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? Adv. Neural Inf. Process. Syst., vol. 30, Cu...

work page doi:10.1016/j.neunet.2025.108161 2026
[10]

On Calibration of Modern Neural Networks

Guo C, Pleiss G, Sun Y, Weinberger KQ. On Calibration of Modern Neural Networks. Proc. 34th Int. Conf. Mach. Learn., PMLR; 2017, p. 1321–30. [99] Lakshminarayanan B, Pritzel A, Blundell C. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. Adv. Neural Inf. Process. Syst., vol. 30, Curran Associates, Inc.; 2017. [100] Geifman Y, El...

work page doi:10.1177/1754073912468165 2017