nucleation

LIT-1484: The Hive Mind is a Single Reinforcement Learning Agent

2 min read

Properties4
StatusActive — read in full 2026-10-24 (NOTE-1132) from arXiv v6 (21 July 2026, CC BY 4.0); worth reading as a proved aggregation result: in an infinite, well-mixed population, two imitation rules with no individual learning have the expected update of a single bandit-learning agent. Imitation of success gives Cross Learning (the Taylor replicator dynamic), the honey-bee weighted voter model gives what the authors call Maynard-Cross Learning (the Maynard-Smith replicator dynamic). The macro-agent is a stateless bandit policy over the members' types. The biological model is deliberately partial, the simulations qualitative, and pure imitation has no way back once a type is lost or the environment changes.
Tagscomplex-systems, learning-theory, game-theory, agency, anthology-candidate
Cited as “paper” byNOTE-1132 — Soma et al. — The hive mind is a single reinforcement learning agent
Cited as “source” byTHEORY-899 — In an infinite, well-mixed population whose members each sample the quality of their own type, imitation of success has the expected update of a single Cross Learning bandit agent (the Taylor replicator dynamic), and the weighted voter model, in which members broadcast their type in proportion to their sample and adopt the first type they perceive, has the expected update of Maynard-Cross Learning, π_a ← π_a + (r_k / v^π)(1[a = k] − π_a), the Maynard-Smith replicator dynamic; the population vector is the policy and no member learns, THEORY-905 — In simulations of the two imitation protocols, populations small enough to be finite often settle on a suboptimal option because pure imitation can remove a type entirely, and once settled the protocols give no way to re-explore when the environment changes; variants in which members compare quality estimates converge faster than blind weighted voting, at the cost of more communication and comparison

Graph View

Soma, Bouteiller, Hamann & Beltrame (2024; Swarm Intelligence 2026) — arXiv:2410.17517

Key takeaways

  • The equivalence. Imitation of success aggregates into the Taylor replicator dynamic, which is the expected Cross Learning update (Proposition 1, which the authors call a consequence of known results). The weighted voter model aggregates into the Maynard-Smith replicator dynamic, the expected update of Maynard-Cross Learning (Proposition 2, the main result). The population vector and the policy are the same point on the simplex (THEORY-899).
  • No member learns. In the weighted voter model a bee broadcasts its type in proportion to its own quality sample and adopts the first type it perceives. “In the weighted voter model the individual bee itself is not modeled as an RL agent: it only “blindly” imitates its peers, irrespective of their payoffs!” (p. 10).
  • Members are parallel samples. α = 1/N plays the part of a learning rate, so a large swarm is a bandit learner evaluating its policy in N environments at once.
  • Limits. The model omits spontaneous commitment, stop signals, direct switches and quorum sensing. Simulations are “interpreted qualitatively rather than quantitatively” (p. 13). Finite populations can lose a type entirely, and pure imitation gives no way to re-explore after the environment changes (THEORY-905).

Why this is here

Read at the owner’s request on 2026-10-24. It is a demonstrated case of an organized whole that learns as one agent with no learning member, which bears on the distributed-agency line, on THEORY-043’s mind of mindless agents, and on CLAIM-443’s learned prior (NOTE-1132). It carries anthology-candidate because the authors offer it for designing collective RL systems. It carries no instruction for ML practice.

Backlinks