cs.NEApr 8, 2026

The Principle of Maximum Heterogeneity Optimises Productivity in Distributed Production Systems Across Biology, Economics, and Computing

Authors: Guillhem Artis, Danyal Akarca, Jascha Achterberg

Organizations: Callosum, London, United Kingdom (www.callosum.com) · Imperial College London, London, United Kingdom · University of Cambridge, Cambridge, United Kingdom · University of Oxford, Oxford, United Kingdom

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The world is full of systems of distributed agents, collaborating and competing in complex ways: firms and workers specialise within economies, neurons adapt their tuning across brain circuits, and species compete and coexist within ecosystems. In that context, individual research fields built theories explaining how comparative advantage drives trade specialisation, how balanced neural representations emerge from sensory coding, and how biodiversity sustains ecological productivity. Here we propose that many of these well-understood findings across fields can be captured in one simple joint cross-disciplinary model, which we call the Distributed Production System. It captures how agent heterogeneity, resource constraints, communication topology, and task structure jointly determine the productivity, efficiency, and robustness of distributed systems across biology, economics, neuroscience, and computing. This model reveals that a small set of underlying laws generates the complex dynamics observed across fields. These can be summarised in our Principle of Maximum Heterogeneity: any distributed production system optimising for performance will converge on an increasingly heterogeneous configuration; environmental demands place an upper bound on the degree of heterogeneity required; and the communication topology determines the spatial scale over which heterogeneity spreads, with this principle applying recursively across all layers of nested production systems. Beyond explaining existing systems, these principles act as a blueprint for constructing ideal ones. We demonstrate this by suggesting specific redesigns for compute systems executing large-scale AI. In total, The Principle of Maximum Heterogeneity reveals a unique convergence of complex phenomena across fields onto simple underlying design principles with important predictive value for future distributed production systems.

Figures & tables

Figure 1: Intuitive visualisation of the distributed production system model in which each agent contributes an individual skill density and connected agents jointly produce an output that can match or fail to match a given demand. The lower half of the figure shows three agents ( w1w_{1} , w2w_{2} , w3w_{3} ), each characterised by an individual skill density concentrated around a specialism. The interaction network Q(1)Q_{(1)} encodes which agents communicate. Together, skill densities and network structure determine the joint production function shown in the upper half, computed as W(1)=(Q(1)+I) w(1)⋅1⃗W_{(1)}=(Q_{(1)}+\mathbf{I})\,w_{(1)}\cdot\vec{1} where 1⃗\vec{1} denotes the vector full of ones. The right panel illustrates that the same production function can be well-suited to one demand shape (Demand A) and poorly-suited to another (Demand B), motivating the optimisation problem at the core of the model.
Figure 2: Intuitive visualisation of the distributed production system model with constraints. The resource constraint acts as a parameter on the production function. The communication cost, combined with the network structure, influences the gradient descent by adding a penalty. The second-order constraint, combined with the agents’ skill densities, takes the form of a penalty in the optimisation loss. The arrows indicate the effects: white denotes the parameters of the model, black the effect of the constraints and grey the combined effects of the constraints and model parameters.
Figure 3: According to the theory of adaptive radiations, as species proliferate in a new environment, they become increasingly specialised. For two workload shapes (uniform, multi-Gaussian) we study the average specialisation of the converged model as we increase the number of agents. The predicted trend according to the theory of adaptive radiations is highlighted with the dotted line. As expected, the multi-Gaussian workload with its ‘ecological niches’ within its distribution replicates the expected trend.
Figure 4: Homogeneous and heterogeneous temporal biomass production across unimodal environmental pressures. Top: Production time series for the single drifting Gaussian task. The heterogeneous community (green) maintains higher and more stable production than the homogeneous community (blue). Dashed lines indicate temporal means. Bottom: Per-task mean production (left) and coefficient of variation (right) for the sixteen environmental pressures. The heterogeneous community (green) exceeds the homogeneous one (blue) in mean production and falls below it in CV for every task ( p<10−4p<10^{-4} , paired tests).
StatisticHomogeneousHeterogeneousTestpp -valueDirection
Mean production0.0210.0210.049  (+134%)0.049\;(+134\%)t=7.24t=7.242.86 ⁣× ⁣10−62.86\!\times\!10^{-6}16/1616/16
Variability a1.171.170.52  (−56%)0.52\;(-56\%)t=8.06t=8.067.91 ⁣× ⁣10−77.91\!\times\!10^{-7}16/1616/16
Table 1: Statistical comparison of community biomass production and variability. Summary statistics, paired tt -test results, and effect directions comparing homogeneous and heterogeneous communities across 16 unimodal environmental pressures.
Figure 5: Agent populations naturally adapt their filters to match local stimulus distributions. Two independent pairs of agents (populations) are optimised on different stimulus distributions, mirroring distinct regions of a visual field. Solid purple lines represent the expected stimulus (workload), and dashed blue lines represent the learned filters. (Left, Center) When isolated and presented with different expected stimuli, the two populations independently converge to specialised heterogeneities that perfectly match their local workloads. (Right) The combined full system demonstrates how these spatially segregated, specialised populations collectively cover the broader, bimodal distribution of the entire environment, analogous to the heterogeneous distribution of colour filters across the retina.
Figure 6: Network topology underlying the MD system experiment. To run the MD system experiment we use a star-like topology, where we expect the central agent to converge on a generalist role, with peripheral agents converging on specialised roles.
Figure 7: The model converges on brain-like structure-function topology. In the brain we observe that the topologically central region (MD system; 0) takes a generalist role during problem solving, with more peripheral regions taking more specialised roles (1-8). We see these findings replicated in our model across the skill distributions (left) and their respective specialisation levels averaged across multiple optimisations from random initialisation (right). Error bars on figures represent standard errors.
Figure 8: Three-tier network topology. The core agent (green) connects to all penumbra agents (blue), each of which links to three peripheral agents (purple) that have no other connections.
Figure 9: Specialisation increases from core to periphery. Mean specialisation per tier averaged over 100 simulations; vertical bars denote one standard error below and above the mean. Error bars on figures represent standard errors
Figure 10: Production profiles of two countries. Top row: autarchy. Bottom row: trade. Under trade, each country specialises toward a distinct demand peak, yielding higher combined coverage.
Figure 11: Trading between countries increase realised consumption. Realised consumption (shaded area) under autarchy (left) and trade (right). The integral of min⁡(d,W)\min(d,W) is strictly larger under trade, confirming collective gains from specialisation and exchange.
Figure 12: Productivity is positively correlated with heterogeneity and specialisation. Monte Carlo validation ( 5050 random tasks). Distribution of Spearman rank correlations between productivity and heterogeneity (left) or specialisation (right) across a stiffness sweep. Vertical lines indicate the sample mean; both are significantly positive.
Figure 13: Labour market efficiency determines the specialisation of hired co-workers. Specialisation of the optimal hire as a function of labour market inefficiency α\alpha (right-to-left: decreasing friction). Three regimes are visible: a high-specialisation plateau under efficient markets, a transition region of declining specialisation as friction increases, and a low-specialisation plateau under maximally inefficient markets where search cost outweighs any complementarity gain.
Figure 14: Per-market-evolution mean return (left) and coefficient of variation (right) for the homogeneous portfolio and the diversified portfolio across 1616 unimodal market conditions spanning narrow to broad sector spectra. The diversified portfolio achieves higher mean return and lower CV across all evolutions.
MetricTestStatisticpp -value
Mean returnPaired ttt=7.31t=7.312.6×10−62.6\times 10^{-6}
Coeff. of variation 1Paired ttt=8.00t=8.008.7×10−78.7\times 10^{-7}
Table 2: Statistical tests across different market evolutions.
Figure 15: Learning performance evolution with heterogeneity and resource. (Left) Learning performance as a function of heterogeneity under varying stiffness. Each point corresponds to a single stiffness value; the regression line ( R2=0.92R^{2}=0.92 ) confirms a strong positive association. (Right) Learning performance of the homogeneous network ( s=20s=20 , H≈0\mathcal{H}\approx 0 ) as resource constraints are relaxed. Recovering high performance requires a large increase in effective capacity. For readability, the resource constraints on the x-axis correspond to 8 times the resource constraints used in simulation.
Figure 16: Scaling behaviour of the distributed production model. The model recovers multi-parameter scaling laws (left) while demonstrating the efficiency trade-offs of homogeneous scaling (right). (Left) Recovery of Chinchilla-type scaling. L=A/Nα+B/Dβ+L0L=A/N^{\alpha}+B/D^{\beta}+L_{0} fit to (N,D,L(N,D))(N,D,L(N,D)) . (Right) One-parameter scaling. L=B/Dβ+L0L=B/D^{\beta}+L_{0} fit to (D,L(D))(D,L(D)) , showing overproduction trade-offs.
Figure 17: Performance and robustness matrices for the homogeneous and heterogeneous models across two tasks. Left: initial performance. Centre: performance after a π/4\pi/4 rotation of the task distribution. Right: robustness index ρ\rho (lower is better). The heterogeneous model exhibits a lower mean robustness index, indicating greater stability under distributional shift.
Figure 18: Heterogeneity and specialisation scale with the system size. The line shows the mean value for a system with a specific number of agents averaged across the 8 workloads studied. Shading represents one standard-deviation around the mean.
Figure 19: Logistic scaling of heterogeneity with system size. The dashed line is the fitted scaling relationship between heterogeneity and logarithm of system size. Shading represents one standard-deviation around the mean.
Figure 20: Heterogeneity and specialisation scale with the system size. Each dot is the mean value for a system with a specific number of agents averaged across the 16 workloads studied.
Figure 21: The low-level heterogeneity of the workload has the highest correlation with the optimal system heterogeneity. Pearson correlations between 12 workload shape metrics and the asymptotic system heterogeneity H∞\mathcal{H}_{\infty} , computed across 8 workload profiles (1 to 8 evenly spaced Gaussian peaks). Positive correlations (blue) indicate features that increase with H∞\mathcal{H}_{\infty} ; negative correlations (purple) indicate features that decrease. Four features pass Bonferroni correction ( α=0.004\alpha=0.004 ): Hlow\mathcal{H}_{\text{low}} , peak height, circular variance, and entropy.
Figure 22: The two strongest univariate predictors of system heterogeneity. Each dot is one workload profile, coloured by its peak family (mono through oct). (Left) Low-level heterogeneity Hlow\mathcal{H}_{\text{low}} versus asymptotic system heterogeneity H∞\mathcal{H}_{\infty} ( R2=0.81R^{2}=0.81 ). (Right) Peak height versus H∞\mathcal{H}_{\infty} ( R2=0.79R^{2}=0.79 , negative relationship). Regression lines show the best linear fit. Together these two features confirm that workloads which are themselves more heterogeneous (high Hlow\mathcal{H}_{\text{low}} , low peak height) drive the system towards higher heterogeneity.
Figure 23: Effect of the maximum communication component size. Left: heterogeneity as a function of kk . System heterogeneity (blue) and within-component heterogeneity (green) both increase monotonically with kk ; at k=8k=8 they merge since a single component spans the entire system. Shaded bands show [q10,q90][q_{10},q_{90}] over 20 seeds. Right: Optimisation loss vs. kk . The cost drops by nearly four orders of magnitude from isolated agents ( k=1k=1 ) to full connectivity ( k=8k=8 ), showing the large efficiency gain from collaboration.
Figure 24: Optimised agent positions for each component size. Each dot is one agent from one of 20 Monte-Carlo seeds; vertical jitter separates overlapping points. Dotted lines mark the two task peaks. At k=2k=2 agents are confined near the peaks; larger components enable a broader spread across the task support.
Figure 25: Agent positions coloured by specialisation level. Left: each dot is one agent, placed at its optimised μ\mu ; colour intensity encodes breadth σ\sigma (pale == generalist, dark purple == specialist). Right: distribution of 1/σ1/\sigma for each component size; diamonds mark the median. Larger components enable agents to become simultaneously more spread out and more specialised.
Figure 26: Locality of heterogeneity induced by communication costs. Plotting within- and between-cluster distances across varying communication costs reveals sharp phase transitions from maximal system heterogeneity, to local homogenisation within tightly connected clusters, and ultimately to complete global homogenisation.
Figure 27: The efficient system-level architecture is a heterogeneous architecture. When forcing agents to find the efficient solution by imposing pressure on their resources, the system increasingly converges upon a heterogeneous setup, across workloads. The shaded area represents one standard deviation around the mean.
Figure 28: Performance and heterogeneity decrease concurrently with stiffness. Aggregated across varying resource constraints, the system’s performance and heterogeneity exhibit joint degradation in response to applied stiffness.
Figure 29: Summary robustness for the catalogue suited to the homogeneous system . Despite lower mean production, the heterogeneous system exhibits significantly lower CV and higher minimum production under extreme events.
Figure 30: Workload-production comparison across different hardware instantiations. Workload is represented in blue and production functions are coloured in purple. Top-left, homogeneous hardware: The production function is narrow and poorly matched to the workload (perf. ≈0.69\approx 0.69 ). Top-right, no hardware constraint: agents freely specialise and nearly perfectly cover the workload (perf. ≈0.99\approx 0.99 ). Bottom-left, perfectly aligned heterogeneous hardware:strong constraints are present but aligned with the optimal specialisation, largely restoring performance (perf. ≈0.94\approx 0.94 ). Bottom-right, quasi-perfectly aligned heterogeneous hardware: small misalignment degrades performance but remains substantially better than the homogeneous case (perf. ≈0.81\approx 0.81 ).
Figure 31: Summary of system performance across hardware configurations. Performance denotes the full objective 1/(1+L)1/(1+\mathcal{L}) , task performance isolates the workload-coverage component 1/(1+Lm)1/(1+\mathcal{L}_{m}) , excluding the hardware mismatch penalty. Aligned heterogeneous hardware nearly matches the task performance of the unconstrained case, while homogeneous hardware suffers a large deficit. Misalignment degrades both metrics but remains preferable to homogeneity.
Figure 32: Heterogeneous hardware reaches peak productivity at smaller system sizes. Smoothed mean productivity ( ±1\pm 1 SE, 20 runs) versus system size NN . The heterogeneous system (blue) saturates near its maximum already at N≈4N\approx 4 , whereas the homogeneous system (purple) requires roughly N≈11N\approx 11 to reach the same level.
Figure 33: Heterogeneous hardware is more performant than homogeneous hardware with performance meeting in the limit of infinite resources. The purple curve represents homogeneous scaling and the green curve heterogeneous scaling. Figure shows a schematic and does not show an actual analysis based on model simulations.
Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 34: Comprehensive overview of tasks used in the model: (a) general principles, (b) multi-Gaussian scaling, and (c) robustness testing.
Figure 35: Overview of workloads used for extracting the efficiency principle.
Figure 36: Comparison of specialisation levels under different agent weighting schemes.
Figure 37: Specialisation deviations with deviations of agents’ weights.
RegimeE2/E1E_{2}/E_{1}FFpp -value
Low RR (abundant)1.087.210.009
Mid RR1.077.610.007
High RR (scarce)2.3199.81<0.001<0.001
Appendix
Table 3: Heterogeneity carries independent explanatory power for performance under resource constraints. E2/E1E_{2}/E_{1} is the ratio of prediction errors of the restricted model ( M2M_{2} , resource constraints only) to the full model ( M1M_{1} , resource constraints and heterogeneity), and FF is the corresponding FF -statistic. While heterogeneity provides a marginal predictive benefit under resource abundance, its explanatory power for performance becomes dominant and highly significant as resources grow scarce.
Figure 38: Example production time series for the unimodal catalogue (task 4) under wave, Brownian motion, and extreme-event regimes. Blue: homogeneous system; purple: heterogeneous system.
Figure 39: Summary robustness statistics (mean, CV, minimum) averaged across all 1616 unimodal tasks. Hatched bars: homogeneous; solid bars: heterogeneous. Colours indicate regime (blue: wave; purple: BM; dark blue: extreme).
Figure 40: Example production time series for the diverse catalogue (task 4) under wave, Brownian motion, and extreme-event regimes.
Figure 41: Summary robustness statistics for the diverse catalogue ( 1616 tasks spanning uniform through quintmodal configurations).
Figure 42: Example production time series for the homo-suited catalogue (task 4). The homogeneous system starts with high production (task initially at μ=π\mu=\pi ) but suffers large swings; the heterogeneous system is steadier.
Figure 43: Summary robustness for the homo-suited catalogue. Despite lower mean production, the heterogeneous system exhibits significantly lower CV and higher minimum production under extreme events.
More in this topic

Multi-Agent System Optimization

Multi-Agent System Optimization studies how to improve agent teams’ coordination, performance, resource use, and reliability.

See all 160 papers

Explore similar work

CardsList
Aug 19, 2026

Biological-Hybrid Intelligence: A Conceptual Framework for Distributed Biological--Artificial Computation

Biological and artificial systems offer complementary forms of adaptation, learning, and computation, with advances in in-vitro neurotechnology increasingly enabling bidirectional coupling between them. As these systems become more tightly integrated, a key architectural question is how task-relevant computation should be distributed across both substrates. Yet existing biohybrid solutions optimise the biological substrate, the AI model, or their interface without explicitly addressing how such computation is allocated, reassigned, and evaluated. We introduce Biological-Hybrid Intelligence (BHI), a conceptual framework for distributing computation across adaptive biological and artificial substrates coupled through a bioelectronic interface and coordinated by an orchestrator. BHI treats both substrates as computational entities whose computational responsibilities may change during operation. BHI requires reciprocal co-adaptation and differs from systems that merely decode biological activity, stimulate a living substrate, or adapt a single component. BHI further defines three operating modes: adversarial, collaborative, and codependent, distinguished by whether the substrates compete, divide computational labour, or become mutually necessary for task performance. BHI provides a common basis for comparing computational frameworks, defining benchmarks for latency, viability, interface bandwidth, learning efficiency, and reproducibility. It also highlights governance considerations arising from reciprocal stimulation, adaptation, and data exchange. More broadly, BHI invites computer scientists to consider biological substrates as active computational resources and to ask not only how a task should be computed, but where its computation should reside. BHI therefore reframes biological-artificial integration as a system-level problem of computational allocation, coordination, and control.

Michael Taynnan Barros, Sergio Lopez Bernal, Reinhold Scherer

School of Computer Science and Electronic Engineering, University of Essex, Colchester, UK · Department of Information and Communications Engineering, University of Murcia, Murcia, Spain

Adaptive ComputationHeterogeneous Computing
AskPDF
Apr 29, 2026

Split over nn resource sharing problem: Are fewer capable agents better than many simpler ones?

In multi-agent systems, should limited resources be concentrated into a few capable agents or distributed among many simpler ones? This work formulates the split over nn resource sharing problem where a group of nn agents equally shares a common resource (e.g., monetary budget, computational resources, physical size). We present a case study in multi-agent coverage where the area of the disk-shaped footprint of agents scales as 1/n1/n. A formal analysis reveals that the initial coverage rate grows with nn. However, if the speed of agents decreases proportionally with their radii, groups of all sizes perform equally well, whereas if it decreases proportionally with their footprints, a single agent performs best. We also present computer simulations in which resource splitting increases the failure rates of individual agents. The models and findings help identify optimal distributiveness levels and inform the design of multi-agent systems under resource constraints.

Karthik Soma, Mohamed S. Talamali, Genki Miyauchi +3

MIST lab, École Polytechnique de Montréal, Montreal, Canada · School of Electrical and Electronic Engineering, The University of Sheffield, Sheffield, UK · Department of Computer and Information Science, University of Konstanz, Konstanz, Germany +1

Multi-Agent System OptimizationMulti-Agent SystemsMulti-Agent Resource Allocation
AskPDF
Jun 18, 2026

Optimal Order of Multi-Agent and General Many-Body Systems

This paper develops a general framework for analyzing multi-agent systems with feedback loops between agents actions and collective observations. The framework is built on two fundamental agent-level variables: power, which measures agent influence on collective outcomes, and response functions, which determine how agents react to observations. We derive how macroscopic properties, including total power, useful power, entropy, order, fragility, and mobility, emerge from these two variables of heterogeneous agents. To study the trade off between growth and resilience, we introduce a system-level utility function parameterized by a risk-appetite coefficient and derive an optimal degree of order that balances productivity, stability, and adaptability. The analysis suggests that stronger synchronization can increase collective output but may also increase systemic fragility and reduce mobility. We further argue that order, entropy, information, and useful energy are task-dependent and system-relative concepts whose meanings depend on the objectives of the system. By measuring and designing agent power distributions and response functions, it may be possible to better understand, predict, and optimize collective behavior and identify the conditions under which collective intelligence and optimal order emerge.

Jake J. Xia

Harvard Management Company · Research Affiliate at Massachusetts Institute of Technology

Dynamical SystemsMulti-Agent System OptimizationMulti-Agent CoordinationMulti-Agent SystemsEmergent Behavior in Multi-Agent Systems
AskPDF