When does more communication spread error instead of truth?

Communication topology can change collective accuracy even when the agents, evidence process, population size and total edge count are held fixed. Higher degree can speed convergence, but faster agreement does not guarantee truth. The model suggests an epistemic-firebreak hypothesis for organizations; it does not test one.

Michael Crosscombe and Jonathan Lawry · Distributed Autonomous Robotic Systems: 15th International Symposium, DARS 2021, Springer Proceedings in Advanced Robotics 22, pages 82-94 · January 3, 2022 prototype 5 min read Explained by Superalignment Research

Read the original

Topics: agents, theory, collective learning, network science, organizational design, social epistemology

The 60-second answer

AI assurance depends on communication across models and teams, but extra links can transmit shared error through the same channels that transmit useful evidence.

Crosscombe and Lawry show a speed-accuracy tradeoff in a stylized collective-learning model. Their regular-network experiment shows that higher-degree networks often converge faster. In a separate rewiring experiment with total edge count and nominal k fixed, random shortcuts can leave more final error under noisy, sparse evidence. In one reported cell with degree 10, evidence rate .01 and evidence error .2, final average error rises from .027 in a regular network to .085 with rewiring probability .1 and .162 in a random network. The result changes across the parameter grid and does not establish that sparse or regular networks are always better.

The simulation holds agent rules and evidence fixed while changing who can exchange beliefs with whom.
Higher degree often speeds convergence in regular networks, but convergence speed is not the same as final accuracy.
In one sparse, noisy evidence cell, rewiring raises final average error from 0.027 to 0.162 while total edge count stays fixed.
The result is a mechanism study, not evidence that real organizations should communicate less.

Written for: Technical generalists who can read a simulation result without network-science training. Useful prerequisites: A graph with agents as nodes and communication links as edges, Noisy evidence, The difference between agreement and accuracy.

The question: How does the topology of an agent communication network change the speed and accuracy of collective learning?
What the authors did: The authors simulate 100 agents learning 100 binary propositions. Agents receive noisy evidence, exchange beliefs with one neighbor along a small-world network and fuse those beliefs with a fixed rule. Across 100 runs per parameter cell, they vary network degree, rewiring probability, evidence rate and evidence error, then measure convergence time and final average error.
The source: The Impact of Network Connectivity on Collective Learning

When random shortcuts leave more final error

Regular network selected. Average final error is 0.027.

Exact values read from Figure 5c

Rewiring probability	Average final error	What remains fixed
rho = 0	0.027	100 agents, total edge count implied by k = 10, r = .01, evidence error = .2, 100 runs per cell. Rewiring may change individual node degrees.
rho = .1	0.085
rho = 1	0.162

Exact values from Figure 5c at nominal k = 10, evidence rate .01 and evidence error .2, with 100 runs per parameter cell. Average final error is .027 at rho 0, .085 at rho .1 and .162 at rho 1. The paper describes .162 as a 600 percent increase over .027. Numerically, it is six times the value, which is a 500 percent relative increase. The control selects reported cells and does not imply a continuous curve. Bar lengths use a zero baseline and the reported Figure 5c values. The three bars are discrete simulation cells, not a fitted function over rewiring probability.

Walk through the argument

Each step below points to the source location that carries it. The source map follows the walkthrough for source-level checking.

Put every agent in the same controlled world

The model contains 100 agents learning 100 binary propositions. Agents receive noisy evidence and combine it with beliefs received from a neighbor using one fixed fusion rule. There are no strategic incentives, status differences or organizational roles.

That simplicity is a strength for mechanism testing. When collective error changes, the simulation can attribute the change to network and evidence parameters rather than to a different kind of agent.

Source: Section 3, A propositional model for collective learning, Section 5, Agent-based simulations

Vary degree and rewiring separately

Degree controls how many local links a regular network starts with. Rewiring replaces eligible links with shortcuts. In the rewiring experiment, total edge count and nominal degree stay fixed, although individual node degrees can change.

This distinction matters. More links and different links are separate interventions. A claim about faster convergence under higher degree cannot be silently reused as a claim about accuracy under random rewiring.

Source: Section 4, Small-world networks, Section 5.1, Convergence results for regular small-world networks

Track both time to agreement and error

The study measures convergence time and final average error across 100 runs for each parameter cell. Regular networks with higher degree often converge faster because beliefs can travel through more routes.

Fast convergence is operationally attractive, but it does not reveal whether the settled belief is correct. The second measure is essential because the same connection that spreads evidence can also spread a mistaken belief.

Source: Section 5, Agent-based simulations, Section 5.1, Convergence results for regular small-world networks

Hold the evidence regime fixed and rewire

Figure 5c fixes nominal degree at 10, evidence rate at 0.01 and evidence error at 0.2. Final average error is 0.027 in the regular network, 0.085 at rewiring probability 0.1 and 0.162 in the random network.

The last value is six times the first. These are three discrete simulation cells, not a fitted law. Other cells in the parameter grid move the boundary, so the safe conclusion is conditional.

Source: Section 5.2 and Figure 5c, Section 5, Agent-based simulations

Use the mechanism as a question for organizations

A shortcut can remove an epistemic firebreak by moving useful evidence and shared error together. That suggests a test for assurance systems: are apparently independent teams or models still drawing from meaningfully different information paths?

The paper does not test organizations. Real teams have authority, memory, incentives and unequal expertise. The simulation earns a hypothesis about informational dependence, not a prescription to reduce communication.

Source: Section 6, Discussion and Conclusion, Section 5.2 and Figure 5c

Source map

These are the source locations that carry the argument. Use them to check this explanation against the original rather than trusting the summary alone.

Where the argument lives

Locus	Why it matters	Source
Section 3, A propositional model for collective learning	Defines the agents, binary propositions, noisy evidence and belief-fusion rule used throughout the simulations.	Open source →
Section 4, Small-world networks	Defines degree and rewiring probability, including what rewiring preserves and what it changes.	Open source →
Section 5, Agent-based simulations	Defines the parameter grid, 100-run cells, convergence time and final average error measurements.	Open source →
Section 5.1, Convergence results for regular small-world networks	Reports how higher degree changes convergence speed in regular networks.	Open source →
Section 5.2 and Figure 5c	Reports the highlighted final-error values as rewiring changes under sparse, noisy evidence.	Open source →
Section 6, Discussion and Conclusion	Interprets the speed-accuracy tradeoff and states the simulation's scope.	Open source →

The Assumption Switch

One result. One assumption exposed. Turn it and see what changes.

Assumption under test

Random shortcuts improve information spread enough to offset the loss of regular local structure under noisy, sparse evidence.

Held in the source: With degree 10, evidence rate .01 and evidence error .2, the regular network has average final error .027 across 100 runs per cell.
Turn it: Holding those quantities fixed while rewiring all eligible edges changes the network from regular to random while preserving total edge count and nominal k, although individual node degrees can change.
What changes: Average final error rises to .162 in the reported cell, six times the regular-network value. This is a topology result inside the simulation, not a measured effect in real organizations.

The common misreading

The paper is not evidence that organizations should communicate less. Its strongest lesson is conditional: faster information diffusion can also diffuse error. In the highlighted experiment, rho changes which edges are rewired while preserving total edge count, although individual node degrees can change, and other evidence regimes produce different boundaries.

Outside the ML frame

Organizational design

When does a communication link remove an epistemic firebreak?

Organizations often treat more connectivity as unconditionally good. One plausible mechanism is that a shortcut transmits useful evidence and shared error through the same channel, but the paper does not separately ablate that mechanism. For AI assurance, independent teams, models or evidence paths only add resilience if their information processes remain meaningfully different. That institutional reading is our interpretation, not a tested organizational result.

Where the result stops

The evidence items, binary propositions and belief-fusion rule are simulated. Agents do not have strategic incentives, unequal authority, memory limits, organizational roles or correlated external evidence beyond what the model specifies. Rewiring holds total edge count and the initial k parameter fixed, but individual node degrees can change. The paper explores many parameter combinations, so the three highlighted values are one exact cell rather than a universal ranking of network designs.

The numbers, with their measurands

Each value below states its measurand, evidence type, source location and evidence base when the source reports one. These fields distinguish self reported results from independent measurements.

0.027. average final collective error at rewiring probability 0. Reported as self reported, Figure 5c, k = 10, r = .01, evidence error = .2. Evidence base: 100 simulation runs in the reported parameter cell. Check it →
0.085. average final collective error at rewiring probability .1. Reported as self reported, Figure 5c, k = 10, r = .01, evidence error = .2. Evidence base: 100 simulation runs in the reported parameter cell. Check it →
0.162. average final collective error at rewiring probability 1. Reported as self reported, Figure 5c, k = 10, r = .01, evidence error = .2. Evidence base: 100 simulation runs in the reported parameter cell. Check it →

What remains open

Do the topology effects survive agents with unequal expertise, authority and incentives?
Which network structures preserve independent error correction without paying the full speed cost?
How do correlated evidence sources change the boundary between faster convergence and higher error?
Can real audit organizations measure when a nominally independent channel has become informationally dependent?

What this bears on

Superalignment maintains a public register of the claims it makes and the evidence that would change its mind. The entries below connect this source to the exact public claims it bears on.

C4. Behavioral evaluation cannot carry a deployment decision alone. This record bears on it, indirectly. The simulation holds agent behavior and evidence rules fixed while changing network topology, and group-level error changes materially. This shows why component behavior alone cannot establish a system-level result. See the claim and what would change our mind →

Source audit and review status

What was checked, and against what

Field	Result	Checked	By	Against
`title`	exact	2026-08-17	codex-primary-source-review	source
`authors`	exact	2026-08-17	codex-primary-source-review	source
`date`	exact	2026-08-17	codex-primary-source-review	source
`venue`	exact	2026-08-17	codex-primary-source-review	source
`full_text`	exact	2026-08-17	codex-primary-source-review	source

Review status prototype.
Program collection related prototype; mechanism study; wave 1, release slot unassigned.
Source access public full text, HTML and PDF. Open the reading copy →
Provider-authored safety claim no.
Explained by Superalignment Research.
Reviewed by No named human reviewer yet.
Explainer dates created 2026-08-17; updated 2026-08-17.
AI assistance AI assisted with source discovery, full-text extraction, figure-value checking, first-pass prose and implementation. The page remains a prototype until a named human review is recorded.
Rights and access The arXiv manuscript is openly accessible. The University of Bristol repository and Springer chapter record supply the publication metadata.
Corrections Read the correction policy or report an error.

Provenance

First seen 2026-08-16, via seed_library.py, arxiv shard of the Stampy snapshot; selected for full review in library-cycle-6.
Work id work:network-connectivity-collective-learning, which groups manifestations of the same intellectual work.
Record id arxiv:2106.00655, the natural key for this catalog manifestation.
2026-08-16 seeded from the arXiv shard
2026-08-17 full paper read; publication metadata corrected; unrelated deception tag removed; promoted to an Explained v2 prototype

Full audit data, including this record under id arxiv:2106.00655: /library/records.jsonl. Compact browser index: /library/corpus.json. Catalog method and counts: /library/index.json.