pith. sign in

arxiv: 2606.02411 · v1 · pith:3JQQ3FP5new · submitted 2026-06-01 · ⚛️ physics.soc-ph · nlin.AO

Evolved Collectives Combine Complex Internal Representations with Simple Outputs

Guilherme S. Y. Giardini , John F. Hardy II , Carlo daCunha This is my paper

Pith reviewed 2026-06-28 11:43 UTC · model grok-4.3

classification ⚛️ physics.soc-ph nlin.AO
keywords collective intelligenceswarm behaviorneural controllersevolutionary optimizationhidden layer complexityoutput linearityrequisite complexity
0
0 comments X

The pith

Evolved swarms develop more complex hidden layers while making their output mappings more linear to reach higher collective order.

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

The paper examines evolved agent collectives that use shallow neural controllers under fixed limits on what agents can sense and do. Across thousands of evolutionary runs it reports that the regimes showing the strongest group-level order simultaneously increase the complexity of their internal hidden layers and reduce the nonlinearity of the mapping from those layers to actions. This dual pattern does not appear in the same way in unevolved controls, indicating that optimization is required for the complexity increase while linearization can occur without it. The finding points to a controller architecture that keeps rich internal representations separate from simplified action rules.

Core claim

Under explicit sensory and actuation constraints, the most ordered regimes in evolved swarms exhibit two simultaneous effects: hidden-layer complexity increases while the effective output mapping becomes more linear. Output linearization persists in unevolved controls, whereas the hidden-complexity relation requires the evolutionary process. Behavioral diversity varies nonmonotonically with control parameters according to parameter-specific tradeoffs. These observations are presented as consistent with the law of requisite complexity for internal representations and ecological rationality for outputs, supporting the existence of a partitioned controller organization.

What carries the argument

Partitioned controller organization in shallow neural networks, in which hidden-layer complexity grows while the output mapping linearizes under evolutionary search and explicit constraints.

If this is right

  • Highest collective order arises specifically from the joint increase in hidden complexity and increase in output linearity.
  • Diversity of recurrent collective behaviors is shaped by tradeoffs that depend on the particular control parameters rather than a single optimum.
  • Output linearization can occur without evolutionary adaptation, while hidden-layer complexity increase depends on optimization.
  • Adaptive collective intelligence can emerge from a single controller that maintains both representational richness and action-level simplicity.

Where Pith is reading between the lines

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

  • The same separation of complex internals and linear outputs might appear in other adaptive systems when similar sensory and actuation limits are imposed.
  • Repeating the measurements on controllers optimized by non-evolutionary methods would test whether selection pressure is necessary for the complexity increase.
  • Applying the same analysis to controllers with additional hidden layers could show whether the linearization effect remains stable as depth grows.

Load-bearing premise

The selected measures of hidden-layer complexity and output nonlinearity, together with the evolutionary process across 3024 conditions, correctly capture the link between internal controller structure and collective order.

What would settle it

Finding that the highest-order collective regimes show neither an increase in hidden-layer complexity nor a decrease in output nonlinearity would falsify the reported dual effect.

Figures

Figures reproduced from arXiv: 2606.02411 by Carlo daCunha, Guilherme S. Y. Giardini, John F. Hardy II.

Figure 1
Figure 1. Figure 1: Simplified diagrammatic representation of interacting agents with 3 nearest neighbors. In the diagram, the number of nearest neighbors has been defined as κ = 3 for simplicity. The variables xi and θi are respectively agent i’s position and orientation, the pink sector of a circle represents its field of view. The variables yj are the positions of neighboring agents inside agent i’s field of view, and the … view at source ↗
Figure 2
Figure 2. Figure 2: Classifier-defined phase map over the full 3024-run parameter sweep. Colored points denote simulated parameter combinations labeled by the dominant recurrent regime. The surrounding phase field is obtained by interpolation. Colors encode phase identity rather than numerical magnitude. The reference labels correspond to the following representative regime families: (a) single elongated unpolarized flowing s… view at source ↗
Figure 3
Figure 3. Figure 3: Representative classifier-defined collective regimes: (a) single elongated polarized flow along the major axis, (b) multiple elongated polarized flows along the major axis, (c) single elongated unpolarized stationary state, (d) single round unpolarized stationary state, (e) multiple elongated polarized flows with ambiguous principal-axis propagation, (f) multiple elongated polarized flows along the major a… view at source ↗
Figure 4
Figure 4. Figure 4: Broad and residual association structure among the five principal neural and collective descriptors. Panel A shows the full pairwise correlation pattern. Panel B shows the residual structure after controlling for the remaining variables. Diagonal panels show kernel density estimates of each variable’s marginal distribution. in which different controller components coevolve while contributing differently to… view at source ↗
Figure 5
Figure 5. Figure 5: Combined pairwise comparison matrix for hidden-layer complexity (CLMC), neural nonlinearity (1 − R 2 ), local nematic order (S¯local nem ), and global order (ψglobal ). Lower triangle: run-level scatter plots. Upper triangle: class-averaged pattern summaries. Each panel includes a smoothed trend line and Spearman (S) and Pearson (P) coefficients. Mirrored panels use the same color to indicate the same metr… view at source ↗
Figure 6
Figure 6. Figure 6: Average fitness (⟨φ(t)⟩) and average shared fitness (⟨φsh(t)⟩) as a function of training steps. Geometric Classifier. The classifier assigns compositional labels to each simulation through a hierarchical sequence of geometric tests. First, a Gaussian-blurred occupancy field determines whether the agent distribution is spatially homogeneous or concentrated into localized density peaks. When no peaks are det… view at source ↗
read the original abstract

Collective intelligence emerges from local interactions among agents with limited information, yet how internal controller organization relates to emergent collective order remains unclear. Here, we study evolved swarms with shallow neural controllers under explicit sensory and actuation constraints and compare collective order with hidden-layer complexity and output nonlinearity across 3024 conditions. Under these constraints, the most ordered regimes exhibit two simultaneous and seemingly contrasting effects: hidden-layer complexity increases, while the effective output mapping becomes more linear. The diversity of recurrent collective behaviors varies nonmonotonically across the control parameters, with pattern richness shaped by parameter-specific tradeoffs rather than a single generic constraint optimum. Unevolved controls show that output linearization persists without adaptation, whereas the hidden-complexity relation depends on optimization. These two effects are respectively consistent with the law of requisite complexity and ecological rationality, suggesting that adaptive collective intelligence can arise through a partitioned controller organization in which representational complexity and action-level linearization coexist within the same system.

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.

Referee Report

3 major / 0 minor

Summary. The paper claims that evolved swarms using shallow neural controllers under explicit sensory and actuation constraints exhibit a partitioned organization in the most ordered regimes: hidden-layer complexity increases while the effective output mapping becomes more linear. Across 3024 conditions, behavioral diversity varies nonmonotonically due to parameter-specific tradeoffs. Unevolved controls indicate that output linearization occurs without adaptation while the complexity increase depends on optimization. These effects are interpreted as consistent with the law of requisite complexity and ecological rationality.

Significance. If the operationalizations of complexity and linearity prove robust, the result would be significant for collective intelligence research by showing how representational richness and output simplicity can coexist under constraints, supported by the scale of 3024 conditions and explicit controls separating adaptation effects. This provides a concrete example of partitioned controller organization rather than a generic optimum.

major comments (3)
  1. [Abstract] Abstract: the central claim that ordered regimes simultaneously increase hidden-layer complexity while linearizing the effective output mapping depends on the specific measures of these quantities, yet the abstract (and by extension the reported results) provides no equations, algorithms, or definitions for how complexity is quantified (e.g., number of distinct activations, entropy) or how linearity is assessed (e.g., regression R², mutual information). This is load-bearing because the nonmonotonic diversity pattern and the separation from unevolved controls both rely on these metrics remaining stable and non-circular.
  2. [Abstract] Abstract: no statistical tests, confidence intervals, or details on post-hoc selection are mentioned for the reported patterns across 3024 conditions or the nonmonotonic diversity claim; without these it is impossible to evaluate whether the simultaneous complexity/linearity effects are statistically distinguishable from noise or from metric artifacts.
  3. [Abstract] Abstract: the statements that the two effects are 'respectively consistent with the law of requisite complexity and ecological rationality' are presented as interpretive mappings; without explicit definitions of the fitted quantities or derivations showing that the reported relations do not reduce to a fitted parameter by construction, the consistency claims cannot be assessed as independent evidence.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below, focusing on improving the abstract while noting that detailed operationalizations and analyses appear in the Methods and Results sections of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that ordered regimes simultaneously increase hidden-layer complexity while linearizing the effective output mapping depends on the specific measures of these quantities, yet the abstract (and by extension the reported results) provides no equations, algorithms, or definitions for how complexity is quantified (e.g., number of distinct activations, entropy) or how linearity is assessed (e.g., regression R², mutual information). This is load-bearing because the nonmonotonic diversity pattern and the separation from unevolved controls both rely on these metrics remaining stable and non-circular.

    Authors: We agree the abstract would benefit from greater self-containment. The manuscript defines hidden-layer complexity via the number of distinct activations and entropy over the hidden units, and output linearity via R² from linear regression on the effective input-output mapping together with mutual information. These are applied uniformly across all 3024 conditions and are not circular by construction, as confirmed by the unevolved controls. In revision we will add one sentence to the abstract briefly indicating these operationalizations and directing readers to the Methods for the algorithms. revision: yes

  2. Referee: [Abstract] Abstract: no statistical tests, confidence intervals, or details on post-hoc selection are mentioned for the reported patterns across 3024 conditions or the nonmonotonic diversity claim; without these it is impossible to evaluate whether the simultaneous complexity/linearity effects are statistically distinguishable from noise or from metric artifacts.

    Authors: Statistical support for the reported patterns, including tests for the nonmonotonic diversity and separation from controls, is provided in the Results section. Because abstracts have strict length limits we omitted the full test statistics, but we will revise the abstract to state that the key effects are statistically significant (details in Results) so readers can immediately appreciate that the patterns are distinguishable from noise. revision: partial

  3. Referee: [Abstract] Abstract: the statements that the two effects are 'respectively consistent with the law of requisite complexity and ecological rationality' are presented as interpretive mappings; without explicit definitions of the fitted quantities or derivations showing that the reported relations do not reduce to a fitted parameter by construction, the consistency claims cannot be assessed as independent evidence.

    Authors: The consistency statements are explicitly interpretive, linking the observed partitioned organization (complexity increase plus output linearization) to the qualitative predictions of the two frameworks. The manuscript shows via the unevolved controls that the linearization is not an artifact of the evolutionary fitting procedure, while the complexity increase requires optimization. We will revise the abstract to make this interpretive status and the control-based separation clearer without claiming a formal derivation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical observations are independent of inputs

full rationale

The paper reports results from evolutionary simulations across 3024 conditions, measuring hidden-layer complexity and output nonlinearity against collective order under sensory/actuation constraints. No derivation chain, equations, or fitted parameters are presented in the abstract or described text that reduce a claimed prediction to its own inputs by construction. The consistency statements with requisite complexity and ecological rationality are post-hoc interpretive mappings, not load-bearing derivations or self-definitional steps. No self-citations, ansatzes, or uniqueness theorems are invoked in a way that forces the central claims. The work is self-contained as an empirical comparison, with unevolved controls providing an external benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies no explicit free parameters, axioms, or invented entities; all quantities (complexity measures, linearity metrics, evolutionary fitness) remain undefined.

pith-pipeline@v0.9.1-grok · 5701 in / 1114 out tokens · 19390 ms · 2026-06-28T11:43:48.622355+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references

  1. [1]

    M Ballerini, et al., Interaction ruling animal collective behavior depends on topological rather than metric distance: Evidence from a field study.Proc. Natl. Acad. Sci. U. S. A.105, 1232–1237 (2008)

    2008

  2. [2]

    HFSP J.2, 205–219 (2008)

    I Giardina, Collective behavior in animal groups: Theoretical models and empirical studies. HFSP J.2, 205–219 (2008)

    2008

  3. [3]

    Biol.8, e1002678 (2012)

    J Gautrais, et al., Deciphering interactions in moving animal groups.PLoS Comput. Biol.8, e1002678 (2012)

    2012

  4. [4]

    V Kumar, R De, Efficient flocking: Metric versus topological interactions.R. Soc. Open Sci.8, 202158 (2021)

    2021

  5. [5]

    W Bialek, et al., Statistical mechanics for natural flocks of birds.Proc. Natl. Acad. Sci. U. S. A. 109, 4786–4791 (2012)

    2012

  6. [6]

    HJ Charlesworth, MS Turner, Intrinsically motivated collective motion.Proc. Natl. Acad. Sci. U. S. A.116, 15362–15367 (2019)

    2019

  7. [7]

    Surv.53, 14:1–14:36 (2020)

    S Suran, V Pattanaik, D Draheim, Frameworks for collective intelligence: A systematic literature review.ACM Comput. Surv.53, 14:1–14:36 (2020)

    2020

  8. [8]

    14th Annu

    CW Reynolds, Flocks, herds, and schools: A distributed behavioral model.Proc. 14th Annu. Conf. Comput. Graph. Interact. Tech., SIGGRAPH21, 25–34 (1987)

    1987

  9. [9]

    T Vicsek, A Czirk, E Ben-Jacob, I Cohen, O Shochet, Novel type of phase transition in a system of self-driven particles.Phys. Rev. Lett.75, 1226–1229 (1995)

    1995

  10. [10]

    R Olfati-Saber, RM Murray, Consensus problems in networks of agents with switching topology and time-delays.IEEE Trans. Autom. Control.49, 1520–1533 (2004)

    2004

  11. [11]

    A Costanzo, CK Hemelrijk, Spontaneous emergence of milling (vortex state) in a vicsek-like model.J. Phys. D Appl. Phys.51, 134004 (2018)

    2018

  12. [12]

    IEEE87, 1423–1447 (1999)

    X Y ao, Evolving artificial neural networks.Proc. IEEE87, 1423–1447 (1999)

    1999

  13. [13]

    (MIT Press, Cambridge, MA), (2000)

    S Nolfi, D Floreano,Evolutionary Robotics: The Biology, Intelligence, and Technology of Self-Organizing Machines. (MIT Press, Cambridge, MA), (2000)

    2000

  14. [14]

    Soft Comput.11, 982–990 (2011)

    H Kwasnicka, U Markowska-Kaczmar, M Mikosik, Flocking behaviour in simple ecosystems as a result of artificial evolution.Appl. Soft Comput.11, 982–990 (2011)

    2011

  15. [15]

    O Witkowski, T Ikegami, Emergence of swarming behavior: Foraging agents evolve collective motion based on signaling.PLOS ONE11, e0152756 (2016)

    2016

  16. [16]

    RP Ramos, SM Oliveira, SM Vieira, AL Christensen, Evolving flocking in embodied agents based on local and global application of reynolds’ rules.PLoS One14, e0224376 (2019)

    2019

  17. [17]

    (MIT Press, Cambridge, MA), (1992)

    JH Holland,Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial Intelligence. (MIT Press, Cambridge, MA), (1992)

    1992

  18. [18]

    V Trianni,Evolutionary Swarm Robotics: Evolving Self-Organising Behaviours in Groups of Autonomous Robots, Stud. Comput. Intell. (Springer, Berlin) Vol. 108, (2008)

    2008

  19. [19]

    M Brambilla, E Ferrante, M Birattari, M Dorigo, Swarm robotics: A review from the swarm engineering perspective.Swarm Intell.7, 1–41 (2013)

    2013

  20. [20]

    (2025) OpenReview workshop preprint

    GSY Giardini, CR da Cunha, Fear-driven collective topology: Comparing smart-boid vietoris-rips graphs to animal communication networks via persistent features inNeurIPS 2025 Workshop on AI for Non-Human Animal Communication. (2025) OpenReview workshop preprint

    2025

  21. [21]

    G Cybenko, Approximation by superpositions of a sigmoidal function.Math. Control. Signals, Syst.2, 303–314 (1989)

    1989

  22. [22]

    (Chapman & Hall, London), (1956)

    WR Ashby,An Introduction to Cybernetics. (Chapman & Hall, London), (1956)

    1956

  23. [23]

    P Allen, S Maguire, B McKelvey

    M Boisot, B McKelvey, Complexity and organization–environment relations: Revisiting Ashby’s law of requisite variety inThe SAGE Handbook of Complexity and Management, eds. P Allen, S Maguire, B McKelvey. (SAGE Publications), pp. 279–298 (2011)

    2011

  24. [24]

    (Oxford Univ

    G Gigerenzer, PM Todd, ABC Research Group,Simple Heuristics That Make Us Smart. (Oxford Univ. Press, New Y ork), (1999)

    1999

  25. [25]

    N Wiener, Differential-space.J. Math. Phys.2, 131–174 (1923)

    1923

  26. [26]

    A Genthon, The concept of velocity in the history of brownian motion – from physics to mathematics and vice versa.Eur. Phys. J. H45, 49–105 (2020)

    2020

  27. [27]

    CR daCunha,Introduction to econophysics: Contemporary approaches with Python simulationsed. KWG Ltd. (CRC Press, 6000 Broken Sound Parkway NW, Ste. 300, Boca Raton, FL 33487-2742) Vol. 1, p. 279 (2022)

    2022

  28. [28]

    A391, 2193–2196 (2012)

    A Lipowski, D Lipowska, Roulette-wheel selection via stochastic acceptance.Phys. A391, 2193–2196 (2012)

    2012

  29. [29]

    CR daCunha,Machine learning for the physical sciences: Fundamentals and prototyping with Juliaed. KWG Ltd. (CRC Press, 2385 NW Executive Center Dr., Ste. 320, Boca Raton, FL 33431) Vol. 1, (2023)

    2023

  30. [30]

    (MIT Press, Cambridge, MA), (2023) Paperback reissue; original eBook edition published in 2008

    D Floreano, C Mattiussi,Bio-inspired artificial intelligence: Theories, methods, and technologies. (MIT Press, Cambridge, MA), (2023) Paperback reissue; original eBook edition published in 2008

    2023

  31. [31]

    (Clarendon Press, Oxford) No

    PG de Gennes, J Prost,The Physics of Liquid Crystals, International Series of Monographs on Physics. (Clarendon Press, Oxford) No. 83, 2nd edition, (1993)

    1993

  32. [32]

    (American Mathematical Society, Providence, RI) No

    FRK Chung,Spectral Graph Theory, CBMS Regional Conference Series in Mathematics. (American Mathematical Society, Providence, RI) No. 92, (1997)

    1997

  33. [33]

    R López-Ruiz, HL Mancini, X Calbet, A statistical measure of complexity.Phys. Lett. A209, 321–326 (1995)

    1995

  34. [34]

    (Monticello, IL, USA), pp

    N Tishby, FC Pereira, W Bialek, The information bottleneck method inProceedings of the 37th Annual Allerton Conference on Communication, Control, and Computing. (Monticello, IL, USA), pp. 368–377 (1999)

    1999

  35. [35]

    note on regression and inheritance in the case of two parents.Proc

    K Pearson, Vii. note on regression and inheritance in the case of two parents.Proc. R. Soc. Lond.58, 240–242 (1895)

  36. [36]

    C Spearman, The proof and measurement of association between two things.Am. J. Psychol. 15, 72–101 (1904)

    1904

  37. [37]

    (2026) Accepted; proceedings forthcoming

    CR da Cunha, GSY Giardini, M Turqueti, Llm-based reinforcement learning for controlling robot swarms in2026 7th International Conference on Artificial Intelligence, Robotics, and Control (AIRC 2026). (2026) Accepted; proceedings forthcoming

    2026

  38. [38]

    Rev.63, 129–138 (1956)

    HA Simon, Rational choice and the structure of the environment.Psychol. Rev.63, 129–138 (1956)

    1956

  39. [39]

    Intell.3, 1–19 (2024)

    KJ Friston, et al., Designing ecosystems of intelligence from first principles.Collect. Intell.3, 1–19 (2024)

    2024

  40. [40]

    AK Seth, Distinguishing adaptive from non-adaptive evolution using Ashby’s law of requisite variety inProceedings of the 2002 Congress on Evolutionary Computation (CEC’02). Vol. 2, pp. 1163–1168 (2002)

    2002

  41. [41]

    N Chater, M Oaksford, R Nakisa, M Redington, Fast, frugal, and rational: How rational norms explain behavior.Organ. Behav. Hum. Decis. Process.90, 63–86 (2003)

    2003

  42. [42]

    G Gigerenzer, H Brighton, Homo heuristicus: Why biased minds make better inferences.Top. Cogn. Sci.1, 107–143 (2009)

    2009

  43. [43]

    PM Todd, G Gigerenzer, Environments that make us smart: Ecological rationality.Curr. Dir. Psychol. Sci.16, 167–171 (2007)

    2007

  44. [44]

    SS Iyengar, MR Lepper, When choice is demotivating: Can one desire too much of a good thing?J. Pers. Soc. Psychol.79, 995–1006 (2000)

    2000

  45. [45]

    Stat.7, 1–26 (1979)

    B Efron, Bootstrap methods: Another look at the jackknife.Ann. Stat.7, 1–26 (1979). PNASJune 2, 2026 LaTex v2025 Vol. XXX No. XX eXXXXXXXXXX Preprint manuscript. This version has not been peer reviewed or accepted for publication.9 of 9

    1979