Autonomous Agents · Welfare Theory · Structural Causal Modelling

EHD — Can LLM Agents Care About the World?

Exocentric homeostatic deliberation: anchoring a persistent LLM agent's primary welfare term to an externally monitored, auditable world-state.

Paper

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

Luca Lillo, MSc Data Science and AI, University of Liverpool, 2026

arXiv preprint — under submission

Abstract

This paper introduces exocentric homeostatic deliberation (EHD): a framework in which a persistent LLM agent's primary welfare term is anchored to an externally monitored, auditable world-state rather than to internal comfort variables. The hope term is defined over an interventional distribution under explicit structural-causal modelling assumptions. Pragmatic and epistemic value enter the scoring rule as independently calibrated additive terms. Governance is handled through hard feasibility constraints and a soft penalty term, both inspectable without access to internal operational variables.

Simulations

Reproducible figures from the paper

Proposition 4: Recalibration Convergence

50 independent Robbins-Monro trajectories converging to the true interventional mean μ*(a)=0.35 from initial estimate 0.60.

Proposition 5: EHD vs EFE Ranking Divergence

Actions with equal predicted means but differing variances receive identical EHD pragmatic scores but different EFE KL scores — a restricted existence result.

Section 7: 24-Month Welfare Trajectory

EHD agent operating on an AMR-style monitoring task. Exocentric trigger fires at Wext < θext = 0.45 despite satisfactory internal operational state.

Run simulations locally →

Formal results

Result Statement
Proposition 1 State boundedness
Proposition 2 Monotone passive degradation
Design Principle 1 Governance-transparent attribution
Proposition 4 Robbins-Monro consistency
Proposition 5 Restricted EFE ranking non-equivalence