Swarm Intelligence
Temporary collectives, independent evidence and the cost of agreement
00 / ABSTRACT
A collective is useful when its differences change the decision.
This paper proposes bounded swarms as an optional decision capability inside a controlled AI runtime. A swarm is a temporary set of invocations with distinct roles or evidence paths, a finite shared budget, a synthesis rule and a stopping condition. Its purpose is to resolve a specific uncertainty that a simpler execution path cannot resolve efficiently.
The research question is not whether more model instances can generate more text. It is whether a selected set of perspectives produces a better externally graded decision than a strong single-agent or repeated-sampling baseline at comparable cost. We define the unit of comparison, describe a candidate selection objective and specify evaluation and attribution requirements.
Contributions of this edition
- A bounded swarm contract tied to one decision point.
- A distinction between role diversity, evidence diversity and statistical independence.
- Cost-matched comparisons and leave-one-lens-out attribution.
- 01
Freeze one decision and acceptance.
- 02
Inspect one evidence path.
- 03
Inspect a different failure or source.
- 04
Resolve claims and retain dissent.
- 05
Evaluate the decision outside the vote.
01 / PROBLEM
Agreement can be a correlated mistake.
Multiple instances of the same model may share training data, reasoning habits, prompts and retrieved sources. Renaming them “critic,” “expert” and “planner” does not make their errors independent. A consensus can therefore amplify confidence without adding evidence.
Early sharing creates another problem: later responses may anchor on the first fluent answer. A useful protocol must identify where diversity enters, when communication occurs and what would count as a contribution.
02 / HYPOTHESIS
Spend additional computation at an identified bottleneck.
The hypothesis predicts conditional value, not a universal advantage. Straightforward tasks may be better served by one invocation. A swarm should have a measurable entry condition, such as conflicting sources or a high-impact design choice, and a stop rule when expected information gain no longer justifies cost.
03 / SWARM CONTRACT
One question, one budget, one exit.
| Field | Contract |
|---|---|
| Decision | The exact choice or uncertainty being investigated, with acceptance criteria. |
| Lenses | Responsibilities and source boundaries, selected for a stated reason. |
| Budget | A cap covering branches, synthesis, verification and retries. |
| Contribution | Claim, evidence reference, assumptions, counterexample and unresolved gap. |
| Synthesis | A rule for reconciling incompatible claims without discarding dissent. |
| Exit | Accept, request specific evidence, return a bounded uncertainty, or stop. |
The parent runtime retains authority. A branch may recommend a change; it does not gain permission to execute that change. Independent research can run in parallel, while shared mutations need explicit coordination.
04 / MODEL
Optimize useful coverage, then test the proxy.
S* = arg maxₛ [ Σⱼ wⱼ · coverageⱼ(S) − λ · cost(S) − μ · redundancy(S) ](1)Estimate the terms on development tasks and compare selection against random, fixed-role and similarity-based alternatives. A selector can overfit its own scoring rubric; only held-out task outcomes establish whether the proxy was useful.
The next illustration explains why the count of voices can overstate information. Under equal variance σ² and common pairwise correlation ρ, Var(mean) = σ²[1 + (n − 1)ρ]/n. Solving for an equivalent independent sample size yields the displayed n_eff. This assumption is deliberately simplified.
Five voices. How much independent information?
For five estimates with equal variance and common pairwise correlation ρ, n_eff = 5 / (1 + 4ρ). This is a variance equivalence for a mean, not a prediction of LLM accuracy.
Parameters are reader-selected; no experimental data is used.
05 / PROTOCOL
Separate contribution from persuasion.
- Write the decision brief, source boundary and evaluation rubric before dispatch.
- Collect first-round contributions independently to reduce anchoring.
- Require traceable evidence, material assumptions and a concrete counterexample from each branch.
- Compare claims by evidence and relevance, preserving unresolved contradictions.
- Permit one bounded follow-up only when it addresses a named gap.
- Submit the synthesized artifact to an evaluator outside the agreement process.
For a design review, one lens might examine state consistency while another tests recovery after a partial write. Their outputs should identify incompatible invariants or failure cases. A useful synthesis turns these into a selected design and tests; it does not merely merge prose.
06 / ATTRIBUTION
Did this lens actually change anything?
Track a contribution from its evidence to the final decision. Record whether it introduced a previously absent constraint, invalidated a candidate or caused a test to be added. The number of messages or words is not a credit signal.
Δⱼ = Q(S) − Q(S ∖ {j})(2)Use matched seeds or repeated paired runs where possible. Report confidence intervals and the cost saved by removing a branch. If removing a lens does not reduce outcome quality, its continued inclusion needs another justification.
08 / NEXT EXPERIMENT
Freeze the comparison before seeing the consensus.
| Condition | Purpose |
|---|---|
| A · One strong path | The same model with tools and the full shared budget. |
| B · Repeated sampling | Independent samples with a fixed aggregation rule. |
| C · Fixed lenses | A fixed role set with bounded synthesis. |
| D · Selected lenses | The proposed selector, same total resource cap. |
| Ablations | Remove role differentiation, independent first rounds or provenance requirements. |
Use frozen research and software decisions with adjudicable outcomes. Measure accepted decisions, unsupported claims, missed critical constraints, latency, total cost and the frequency of justified abstention. Grade without revealing the condition. Include simple tasks so routing overhead and unnecessary escalation are visible.
Measure error agreement directly when labels exist. Do not substitute differences in wording for evidence independence. If selected lenses do not beat repeated sampling after costs, the selection mechanism is not supported on that task family.
09 / EVIDENCE SCOPE
An observable substrate is not proof of collective gain.
Rooy is the lab’s computational research substrate for atoms, lenses, gaps and bounded control. Its existence and instrumentation motivate the experiment; they do not establish an improvement over single-agent execution. This paper reports no new swarm benchmark, sample size or effect estimate.
The existing observatory is an experimental artifact and may display a particular state or snapshot. It should not be read as an independently replicated result.
10 / LIMITATIONS
More perspectives create new failure modes.
- Correlated models and shared sources can make apparent diversity illusory.
- Synthesis may discard a correct minority objection or privilege fluent prose.
- Selection may cost more than it saves and may overfit familiar tasks.
- Attribution is difficult when contributions interact.
- Parallel branches can compete for rate limits or shared state.
- No universal optimal swarm size or measured Rooy uplift is claimed.
SOURCES
References & intellectual context
These works provide intellectual context and comparison methods. Their results are not experimental evidence for IO Lab’s hypotheses.
- Wang, X. et al.Self-Consistency Improves Chain of Thought Reasoning in Language Models 2023 · ICLR
- Du, Y. et al.Improving Factuality and Reasoning in Language Models through Multiagent Debate 2023 · arXiv preprint
- Yao, S. et al.ReAct: Synergizing Reasoning and Acting in Language Models 2023 · ICLR
CITE THIS WORK
Citation & edition
Version 0.1 is a research proposal with an explicit protocol. Its analytical illustration is not an empirical result or a performance estimate.
IO Lab. “Swarm Intelligence: Temporary Collectives, Independent Evidence and the Cost of Agreement.” IO Lab Working Paper 002, version 0.1, September 8, 2026. https://io-lab.nglain.com/works/swarm-intelligence
BibTeX
@techreport{iolab2026swarm,
author = {{IO Lab}},
title = {Swarm Intelligence: Temporary Collectives, Independent Evidence and the Cost of Agreement},
institution = {IO Lab},
type = {Working Paper},
number = {002},
year = {2026},
month = {September},
note = {Version 0.1; not peer reviewed},
url = {https://io-lab.nglain.com/works/swarm-intelligence}
}Availability: the full text of this edition is public. The internal repository, raw traces and datasets are not published. No DOI has been assigned.