I would also like to thank prominent figures in the field, Sara Solla, Petra Vertes, @kenmiller.bsky.social, @bendfulcher.bsky.social, @jlizier.bsky.social, @danakarca.bsky.social, MarcusKaiser, Gorka Zamora-López, and Patrick Desrosiers, who provided feedback during lab visits and conferences.

This work has been in progress for 3+ years.
Grateful to my co-authors, Claus Hilgetag, @kayson.bsky.social, and Moein Khajehnejad for their invaluable contributions,

Implications:
🧠 Neuroscience — functional rationale for the evolutionary suppression of strong reciprocal motifs.
🤖 NeuroAI — reciprocity as a tunable design parameter in recurrent & neuromorphic networks.

So why does the brain avoid strong loops?
Because reciprocity systematically hurts computation.

Suppressing strong loops preserves:
working memory
representational diversity
stable but flexible dynamics

We validated this on empirical connectomes (macaque long-distance, macaque visual cortex, marmoset).

Result: the same pattern.
Strong reciprocity consistently undermines memory and representational richness.

Spectral analysis explains why:

- Higher reciprocity → larger spectral radius (instability risk).
- Narrower spectral gap → less dynamical diversity.
- Lower non-normality → weaker transient amplification.

Together → compressed dynamical range.

Interestingly, hierarchical modular networks consistently outperformed random counterparts, but only when reciprocity was low. However, the comparative advantages of network topologies shift with reciprocity, sparsity, and weight distribution

Findings (robust across sizes, densities, architectures):
- Increasing reciprocity (link as well as strength reciprocity) reduces memory capacity.
- Representation becomes less diverse (lower kernel rank).
- Effects are strongest in ultra-sparse networks.

Methods:
- Reservoir computing to isolate structure from learning.
- Networks of 64–256 nodes, in both ultra-sparse and sparse regimes.
- Topologies: small-world, hierarchical modular, core–periphery, hybrid, and nulls.
- Metrics: memory capacity, kernel rank, spectral analysis.

In earlier work (www.biorxiv.org/content/10.1...), we developed Network Reciprocity Control (NRC): algorithms that adjust reciprocity (link + strength) while preserving network structure.

In this study, we apply NRC to systematically test how reciprocity shapes computation in recurrent networks.

The no-strong-loops principle:
Across species (macaque, marmoset, mouse), strong reciprocal (symmetric) connections are rare.

This asymmetry is well known anatomically.
But what are its computational consequences?