SWISwarm Intelligence
SWIBiology atlas
Research snapshot ·

Fish schools / Informed minorities

Where should an expensive or stronger model sit in the swarm?

Collective motion · Couzin model · 5 linked studies · Sources checked: 06.09.2026

What happens in nature?

Couzin and colleagues' numerical model of moving animal groups showed that a few informed individuals could guide the group through local interactions. The evidence here is a controlled model, not a live-fish experiment or LLM comparison.

Transfer boundaryDirectional knowledge in the model is not the same variable as general language-model capability. Measure which knowledge gap a larger model resolves rather than automatically appointing it leader.

From mechanism to protocol

123456Specialist
  1. Producers handling general tasks
  2. Bounded expert help for critical uncertainty
  3. Verification independent of production
SWI adaptation · conceptual communication diagram
In biologyA counterpart to test in your agents
Individual knowing a target directionSpecialist with bounded calls and domain strength
Local interactionConcise guidance and rationale
Coherent group motionIndependent producers sharing acceptance criteria

How can you use it in your swarm?

Spend a limited specialist budget on the task's critical uncertainty.

  1. Locate the knowledge gap

    Producers first generate their own attempt. Tie specialist calls to concrete questions rather than vague requests for help.

  2. Allocate specialist budget

    Make the specialist share of the total budget visible. Count planning, tools and synthesis in that same budget.

  3. Deliver concise guidance

    Ask the specialist for a principle, source and misuse condition. Producers revise their own outputs using this information.

  4. Evaluate the result blindly

    Evaluate the final artifact blinded to specialist use. Retain single larger-model and single smaller-model conditions.

Configure the recipe with your task, agent count and budget

Start with an experiment

Compare a single small model, single large model and small models with bounded specialist calls on the same questions. Report actual aggregate cost.

What happens if you remove the mechanism?
Insert a controlled false assumption into specialist guidance. Measure verifier detection and propagation to other agents.
Primary failure risk
A specialist's wrong answer can pull the team toward one shared error. Do not equate expertise with authority; keep verification independent.

Evidence and related studies

Read biological evidence and agent research separately. The engineering interpretations below are SWI synthesis.

2005 · Journal paperNature 433, 513–516

Effective leadership and decision-making in animal groups on the move

In a numerical model, a small informed minority could guide the group without members recognizing who was informed.

What can I use? Interpretation, limits and provenance

SWI engineering interpretation

Use a specialist as a carrier of critical knowledge rather than a controller of every step.

Limitation

The evidence is a model result, not a universal specialist ratio for fish experiments or LLM teams.

Source record

Iain D. Couzin, Jens Krause, Nigel R. Franks, Simon A. Levin
Publication: 2005-02-03
Abstract review · Checked: 2026-09-06

1987 · Conference paperACM SIGGRAPH 1987

Flocks, Herds, and Schools: A Distributed Behavioral Model

Local motion rules produced collective flock animation without scripting every individual trajectory.

What can I use? Interpretation, limits and provenance

SWI engineering interpretation

Separate collision avoidance, alignment and group cohesion into distinct rules.

Limitation

A behavioral animation model, not a discovery of every interaction rule in real animals.

Source record

Craig W. Reynolds
Publication: 1987 · exact day not verified
Abstract review · Checked: 2026-09-06

2023 · arXiv versionarXiv:2305.14325

Improving Factuality and Reasoning in Language Models through Multiagent Debate

The study examines model instances debating answers over rounds and reports gains on selected tasks.

What can I use? Interpretation, limits and provenance

SWI engineering interpretation

Generate initial proposals independently, then compare rationale and counterevidence.

Limitation

Measure debate cost and shared false assumptions; do not extrapolate to every task class.

Source record

Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, Igor Mordatch
Publication: 2023-05-23
Abstract review · Checked: 2026-09-06

2024 · arXiv versionarXiv:2406.04692 · ICLR 2025

Mixture-of-Agents Enhances Large Language Model Capabilities

A layered architecture uses previous-layer outputs to produce subsequent responses; authors report improvements on selected evaluations.

What can I use? Interpretation, limits and provenance

SWI engineering interpretation

Separate diverse proposers from the role that synthesizes evidence.

Limitation

Response evaluation does not establish equivalent gains in long-horizon tool-using tasks.

Source record

Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, James Zou
Publication: 2024-06-07
Abstract review · Checked: 2026-09-06

2025 · arXiv versionarXiv:2512.08296v3

Towards a Science of Scaling Agent Systems

Version 3 compares 260 configurations across six benchmarks; coordination benefits depend on task structure and communication overhead.

What can I use? Interpretation, limits and provenance

SWI engineering interpretation

Compare single-agent, independent parallel and coordinated teams at the same aggregate budget.

Limitation

Reported counts changed between versions; this entry uses v3. Findings are not a universal scaling law.

Source record

Yubin Kim et al.
Publication: 2025-12-09
Reviewed revision: 2026-04-08
Abstract review · Checked: 2026-09-06