Source
Zollman (2007), LIT-724, read in NOTE-583, for the effect; Rosenstock, Bruner & O’Connor (2017), LIT-733, read in NOTE-559, for its scope. Zollman’s survey (LIT-716) places it among other learning problems.
What was actually shown
In Zollman’s model (LIT-724), agents choose between a known option and a new one that may be better or worse. They update on the outcomes of what they and their neighbours chose. If all of them choose the known option, learning stops for good.
- The effect. In simulated groups of 3–12, a cycle (each agent sees two neighbours) converged on the better option more often than a wheel, and a wheel more often than a complete network. The complete network was much faster. Over every network of up to six agents, density predicted failure and a hub’s centrality did not. The mechanism is that a misleading early run reaches everyone in a dense network at once. In a sparse one, some agents keep trying the better option long enough to recover.
- The scope. Rosenstock et al. (LIT-733) widened the parameters. The effect fell below 2% when the better option’s success rate rose from Zollman’s 0.501 to 0.51, and vanished at 0.525 (10 agents, 1,000 trials a round). It shrank with larger groups and with more data per round. At 100 agents, the complete network was right 99.12% of the time in 20 rounds, and the cycle 100% in 1,977. Zollman’s claim that the ordering holds “for any setting of the parameters” is false. The effect lives where inquiry is hard.
Either result could have come out otherwise: dense networks could have been both faster and more reliable, as they are when agents pool estimates rather than learn by trying (LIT-716).
What this does not say
- Not “less communication is better”. That holds only for this kind of problem (evidence depends on what is chosen), and only when evidence is scarce. For pooling problems denser is faster and no less accurate.
- Not that real scientific or organizational communities are in that regime. The authors of the robustness study call these “how-potentially” models and lower their confidence that the effect is common.
- Not that restricting communication is the remedy. Where the effect holds, Rosenstock et al. prefer stubborn or exploratory agents, or standards for how much evidence a judgment needs, which protect against premature consensus without discarding data.
Connections
- Transient diversity (THEORY-148) carries this further: the benefit comes from diversity of pursuit, which sparse links are one way to keep.
- Longino (THEORY-151). Longino’s norms concern whether criticism has uptake. This account concerns how fast belief spreads. The owner’s essay cites both in §8.2; they answer different questions.
Note of 2026-10-21: an outside audience, and a biased agent
Weatherall, O’Connor and Bruner (LIT-1341), read in full on 2026-10-21 (NOTE-1038), mark a limit of scope this account leaves implicit: the effect concerns the community’s own learning. With an outside audience of policy makers and a propagandist sharing results selectively, the slower consensus that sparse networks give can leave the audience worse off; in their Fig. 5 the complete network is better for the policy makers at p_B = .55 and n = 5 (THEORY-855). And their footnote 14 reports Holman and Bruner’s finding (LIT-751, Unavailable, second-hand) that with a biased agent among the scientists it is best for them to be highly connected. Neither contradicts the account, whose condition is a community of unbiased agents learning for itself. The account is unchanged.
Note of 2026-10-29: what holds it back
Its promote_when is unmet: all of the evidence is simulation of one model family, and no laboratory or field study has been read that varies communication structure in groups learning by trying options and measures accuracy and speed against the difficulty of the problem. It could be narrowed to the result within the bandit models, as THEORY-422 states the wider survey, keeping the wide form as a Proposed THEORY under ADR-047’s kind 1, if a claim comes to need only that.