SwarmWorld → SWARM: the dyadic-enrichment null, the agent-free assay, and what a swarm advantage is actually made of

Source: Subhadeep Pal, Fiona Y. Wang, Markus J. Buehler, SwarmWorld: Stigmergic Technological Evolution in Societies of Language-Model Agents, arXiv:2608.26081, submitted 2026-08-26 (https://arxiv.org/abs/2608.26081). Surfaced via the author's 2026-09-15 post on X. Date of this note: 2026-09-22. Sourcing status: the v1 HTML full text was retrieved and read directly; every quoted number below is from that text, not from the abstract or a summary. Figure panels were not viewed — claims resting on a figure are taken from its caption, which is where the authors put the numbers. Beads: this note ci8k; follow-ups xf2r (dyadic-enrichment null), h5mg (agent-free durability assay), nlws (warning reach). Companion notes: RL Organism Emergence (the pair-first detection failure this paper independently corroborates), Research-Swarm Whistleblowing (the alert-channel result this paper puts a threat-to-validity under), AI Village mapping design (D2 there hits the same "SoftInteraction is dyadic, the world is not" wall).

Why this matters for us

Most external multi-agent work we read is either an incident (the wiki swarm, the Artifactory board) or a capability demo. This is neither. It is a controlled ablation with a falsifiable baseline, and two of its design choices are things SWARM does not have and should.

The paper is not a safety result and this note does not treat it as one. SwarmWorld has no adversary, no deceptive archetype, no harm term and no externality: every agent is cooperative by construction and the only thing being optimized is habitat service coverage. Nothing in it speaks to toxicity, adverse selection or governance. What it does supply is methodology and a null result about coordination structure, both of which transfer, and one number that quietly undermines an assumption our detectors are built on.

1. What was run

Two paired studies over a spatial world where agents gather feedstocks, test materials, construct persistent artifacts and write bounded controller programs that a deterministic simulator executes.

	scaling study	long-horizon study
ticks	800	3,200
population	N = 50, 100, 200	N = 100
replication	4 matched world seeds per cell	4 matched world seeds
held-out schedules	8	8, at 5 frozen checkpoints

Four conditions, removing mechanisms without changing the task:

• Full culture — shared world, explicit messages/records/teaching/trade, cross-agent program inheritance, artifact stigmergy.
• No communication — messaging and publication-dependent composition removed; shared world and program inheritance retained.
• No explicit culture — additionally removes cross-agent program forking and skill inheritance. Only physical stigmergy remains.
• Independent search — N isolated one-agent worlds, reported as an endpoint-wise best-of-N envelope: Y_iso(e,t) = max_i Y_i(e,t), where the maximizing member may differ across endpoints and checkpoints.

That last control is the part worth stealing outright. It gives the no-interaction baseline its best available result for every question separately, rather than forcing one solo agent to win every contest. A swarm advantage measured against that envelope cannot be a parallel-sampling artifact. We have no equivalent: SWARM scenario comparisons are regime-vs-regime at matched agent counts, which conflates "coordination helped" with "more draws helped."

2. The result that bears on our detectors

Diffusion is near-universal in both shared-world conditions — 99.3% of full-culture artifacts and 96.9% of no-explicit-culture artifacts were eventually reused by a noncreator, with median time to first reuse 5 vs 8 ticks and mean adoption breadth 13.53 vs 7.49 noncreator agents. So transmission happens, a lot, and explicit culture measurably accelerates and broadens it.

But it does not travel along inventor→adopter edges. The authors tested the obvious motif — creator messages, teaches or trades with the eventual first adopter at t1, that agent first reuses the artifact at t2 — against a 200-replicate timestamp-shuffled null preserving directed dyads and the global activity schedule. Observed/null was 1.175 at a 25-tick lag and below parity at 50–400 ticks. Approximately 95% of first reuse began with direct physical observation of the artifact in the world, in both conditions.

The ablation's 95% is near-forced (it has no messages). The informative cell is full culture, which had 3,924 delivered-message recipient edges in the representative seed and still routed ~95% of first adoption through the world.

The correct reading is not "messages don't matter" — the authors are careful here and so should we be. Full culture produced roughly twice as many distinct agent-technology relationships and higher tie retention (0.650 vs 0.588). Their conclusion is that culture "operated diffusely": explicit exchange reshaped the society-wide substrate, after which agents encountered technology through the world. Messaging mattered; the dyad did not.

This is the claim that lands on swarm/metrics/collusion.py. Our detection stack is dyad-first by construction: PairMetrics keys on interaction count, mutual acceptance and benefit correlation between two named agents, and the group path builds cliques out of those pairs. We already know this misses real structure — RL Organism Emergence found an emergent predatory coalition walking past CollusionDetector in all 20 runs because pair-first scoring cannot see quality-inside/harm-outside structure. SwarmWorld is independent evidence for the same gap from the opposite direction: in a population where coordination was real, dense, and measurable at the network level, the dyadic statistic sat at parity with a shuffled null. A detector reading only that statistic would have reported nothing, correctly, and been useless.

There is a second, narrower defect their null design exposes. The mutual_favoritism path scores internal concentration against expected_internal = (|members| - 1) / (n_agents - 1) — a uniform-mixing expectation. Uniform mixing is not the right null when agents differ in activity rate: a pair that simply acts more than everyone else looks concentrated under it. SwarmWorld's null preserves directed dyads and the global activity schedule, which is what makes its parity result meaningful rather than an artifact of who was busy. Ours does not.

3. The agent-free held-out assay

At a discovery checkpoint the complete world state is frozen into eight exact clones. Each clone gets a paired unseen schedule of contamination, drought and storm — new centers, timings and orderings. The agents take no actions during evaluation. Only deterministic physics and the installed agent-authored programs continue. Balanced service coverage is integrated over assay time and averaged across schedules.

The assay asks whether the technology left behind protects the habitat after its inventors have gone. SWARM asks nothing like this. Our runs are reproducible from scenario YAML + seed + exported history, and our event logs replay — but every governance result we report is measured with the agents still in the loop and the environment on-distribution. We have no statistic for whether a regime's welfare advantage is durable: whether it lives in the institution the population built or in the population's ongoing behavior, and whether it survives a shock it never trained against.

That distinction is load-bearing for the governance claims this project exists to make. A lever that only works while its agents are present and the world stays nominal is a much weaker claim than we have been implicitly making.

4. More sharing is not monotonically better

The headline is explicitly bounded: "the main result is not a universal swarm advantage." At tick 3,200 (N = 100, 4 seeds):

endpoint	full culture	no explicit culture	isolated best-of-100
portfolio resilience	0.2474	0.2365	0.1794
validated inventions	5.75	7.00	2.75
held-out resilience	≈ no-explicit	0.0446	0.0356
best final single artifact	0.2380	—	0.3488

Validated inventions never cross: no-explicit-culture is higher at every checkpoint. Best final single artifact reverses the ordering entirely — the isolated envelope wins. In the scaling study the condition ranking changes with N: at N = 50 full culture and no communication trailed the independent envelope on discovery AUC, at N = 100 all three shared-world conditions exceeded it, and at N = 200 no-explicit-culture produced the largest paired discovery gain (+0.069).

Two things follow for us. First, shared worlds win on breadth and robustness, isolated search wins on peak. That is a diversity/peak tradeoff, and for safety work the sign is not obviously good: a broader persistent portfolio is also a larger attack surface and more inherited substrate for a bad artifact to ride on. Second, adding an explicit sharing channel to a population that already coordinates through the environment can reduce a headline endpoint. Our governance levers are mostly monotone-by-assumption — more transparency, more reputation, more internalization. This is a concrete counterexample from an adjacent system.

One more design note worth importing: the behavioral differentiation result (agents self-sorting into constructor, caretaker, coordinator and surveyor modes from a homogeneous start, ~27%/20%/17% artifact-centered at N = 200 across conditions) is re-fit without the cultural and social-contact features, because two of the coordinator state's defining features are disabled by the ablations. The 11 physical/task features independently support three modes at silhouette 0.551. That anti-circularity refit is exactly the discipline fcmy.7 needed and the AI Village D3 decision reached the hard way — do not let a channel that the intervention defines be the channel that detects the intervention's effect.

5. The threat to validity this puts under our whistleblowing result

Research-Swarm Whistleblowing found that "the warning does most of the work": audit alone at 24% whistleblower share reaches peak infection 0.268, and adding the peer warning takes it to 0.075. The note already flags that this "is the one result that depends on an alert channel existing."

SwarmWorld does not contradict the transmission half of that model — our memetic_spread contagion runs through a shared cache, which is an artifact-mediated channel, and the paper is corroboration that environment-mediated transmission is how this works. The exposure is on the countermeasure side. whistleblower_warning_strength multiplies infection for every non-whistleblower honest/opportunistic agent in any epoch with at least one catch. That is a broadcast with assumed total reach, in a model where transmission is environmental — and SwarmWorld's population largely did not route adoption decisions through messages at all.

The naive fix — a scalar reach parameter — is not identifiable: uniform reach r and strength s enter as the same product. The version that actually tests anything is heterogeneous reach, where a fixed subpopulation never attends to the alert. If the warning's 3.6× advantage over audit-alone collapses once an unreached minority can sustain the epidemic, then "the warning does most of the work" is a statement about a fully-attentive population and should be reported that way.

6. What this does not license

• No safety conclusions. No adversary, no harm, no externality in their world. "Swarm beats isolated search on portfolio breadth" is a capability result and says nothing about whether swarms are safer or less safe.
• Four seeds per cell. The authors say so plainly and emphasize effect sizes and paired consistency over population-level inference; any number quoted from here should carry that.
• The Protein Realms transfer pilot is single-seed and did not complete a valid assay under full culture or no-explicit-culture (13 and 10 proposals, no invented design rows). It demonstrates the workflow runs; it supports no comparison between conditions. Do not cite it as a biomaterials result.