Published March 28, 2026 | Version v1

Can LLM Agents Care About the World? World-Directed Welfare and Exocentric Homeostatic Deliberation

Authors/Creators

  • 1. ROR icon University of Liverpool

Description

This paper introduces Exocentric Homeostatic Deliberation (EHD), a formal control framework for persistent LLM agents in which the primary welfare signal is defined over an externally monitored, auditable world state rather than over internal comfort variables alone. EHD addresses a gap in the LLM agent literature: whilst existing architectures (ReAct, Reflexion, Generative Agents, cognitive architectures for language agents) equip LLMs with memory, tools, and planning, none provides a formal criterion for detecting welfare deterioration over extended deployments or for triggering renewed intervention on that basis.
The framework combines four elements not jointly present in prior work: (1) a transparent exocentric welfare signal constructed from institutional surveillance data; (2) an intervention-based action-conditional hope term scored via do-calculus rather than correlational conditioning; (3) an additive utility decomposition with separately auditable pragmatic, epistemic, operational, cost, risk, and governance components; and (4) a Robbins–Monro recalibration rule for action efficacy under delayed feedback.
Five propositions and one design principle are established: state boundedness, monotone passive degradation, governance-transparent marginal attribution, mean-square consistency of the mean-update rule, fallback feasibility, and restricted action-ranking divergence relative to a single-step expected free energy (EFE) comparator. An illustrative antimicrobial resistance (AMR) example demonstrates how the epistemic term can reverse local action ranking relative to a pure pragmatic baseline.
The paper sits at the intersection of LLM agent design, active inference, causal decision-making, AI safety and governance, and homeostatic reinforcement learning. It is relevant to researchers working on autonomous agents, agentic AI architectures, world models for language agents, AI alignment, value-sensitive design, multi-session agent control, and the governance of AI-driven decision support systems.
Reproducible simulation code: https://github.com/nepryoon/ehd-simulations
Interactive demo (EHD-Watch): https://www.neuromorphicinference.com

Files

EHD.pdf

Files (497.7 kB)

Name Size
md5:101cd5412b5bce56656572e78de1d68b
497.7 kB Preview Download

Additional details

Software

Repository URL
https://github.com/nepryoon/ehd-simulations
Programming language
Python
Development Status
Active