Contribution
It finds in a honeybee swarm the cross inhibition between competing evidence-accumulating populations that models of neural decision-making assume. Scouts for one nest site send inhibitory stop signals mainly to scouts dancing for the other, during the choice. A model built from the observed behaviour shows why this matters: without it, a swarm choosing between equal sites can stall in a stable deadlock.
Key insight
A collective that chooses by accumulation to a threshold needs a way to suppress the losing alternative, or equal options balance forever. The bees’ stop signal, already known from foraging, does that work when it is aimed at rivals and given often enough. The same structure appears in the Usher–McClelland leaky competing accumulator, so the paper presents the swarm and the brain as two implementations of one decision scheme.
Assumptions
- Nest-site choice is a race to a quorum between populations of scouts, each recruiting by waggle dances (refs. 6, 12).
- A scout that stops dancing for a site soon stops visiting it (refs. 22–23), so inhibition of dancing is inhibition of commitment.
- The model’s equations are mean-field ODEs “rigorously derived from the individual-level interaction rules” observed; their form is in the supplement, not the main text.
- Paint marks did not drive targeting: stop signals were already given to unmarked dancers in the first five swarms.
Key results
- Stop signals in house-hunting. On 2 of 5 swarms, 40 dancers filmed close up; 23 received 109 stop signals (4.7 ± 4.3 per signalled bee) from 40 bees. A signal lasts 0.21 ± 0.10 s; contact is 98% by the head. Signals came mostly in the return phase (96 vs 13 contacts; 73 and 36 expected), and signalled dancers stopped 36 ± 22 s after signals began.
- Signals end dances. On 3 further swarms, 525 signals over 109 dances were concentrated in the last deciles of circuits, also in long dances, which the authors take to show causation, not a common time trend.
- Cross inhibition in the decision phase. Two swarms on an island, two identical nest boxes, scouts painted pink or yellow by box. 98.4% of 1379 signals were attributed to a sender, and 94.5% of 1357 came from marked scouts. Dancers received 213 contra signals (from at least 46 bees) and 24 ipsi (from at least 14), more contra than the marked proportions predict.
- General inhibition in the implementation phase. After worker piping began, contra (185) and ipsi (210) signals were in proportion to the marked scouts.
- One-box control. In the decision phase 26% of 38 dancers received signals with one box, against 66% of 74 with two; in the implementation phase 86% of 148 and 77% of 133, the one-box share being, if anything, higher (P = 0.05).
- Models (Fig. 3, supplement). (i) Direct switching between commitments (ref. 3), with any decay, deadlocks stably between equal sites. (ii) Indiscriminate stop signals deadlock too. (iii) Discriminate stop signals deadlock below a critical signalling rate; above it two stable consensus states appear and the swarm picks one at random; above a critical quality difference it converges on the better site.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Nest-site scouts use the stop signal and it ends dances | strong | 5 swarms; timing of 525 signals over 109 dances |
| C2 | In the decision phase scouts aim stop signals at dancers for the rival site | moderate | 2 swarms, one setup, χ² on pooled signals and on minimum signaller counts |
| C3 | After the choice, signals become general | moderate | same 2 swarms, plus 2 one-box swarms |
| C4 | Discriminate cross inhibition above a threshold breaks deadlock between equal sites | moderate, model | ODE analysis in the supplement, not read here |
| C5 | Swarm and brain converge on one decision scheme because it is reliable | weak | the authors’ “It is tempting to think” |
Concepts
- Stop signal: a vibrational pulse, about 150 ms at about 350 Hz in the foraging literature (0.21 s measured here), usually given by butting the head against a dancer.
- Contra and ipsi signals: from a scout marked for the other box, or for the same box, as the dancer.
- Decision and implementation phases: before and after the onset of worker piping, which marks a quorum reached.
- Stable deadlock: an attractor with equal commitment to two sites.
Connections
The direct-switching model (i) is Marshall et al.’s (2009) optimal-decision model for social insects; the Usher–McClelland model (2001) and Bogacz et al. (2006) supply the neural side. The swarm’s quorum mechanism is Seeley and Visscher’s (2004); the monograph is Honeybee Democracy (2010). None of these is registered.
Bearing on the record
- CASE-145, “Related evidence”. The case cites the paper for collective decisions “with no central controller by cross-inhibition of recruitment”. The paper supports that, with two precisions: the inhibition is aimed at rivals only while the swarm is choosing, and is general once it has chosen; and that it is what breaks deadlock is a model result, holding only above a critical rate of signalling.
- CLAIM-418 and CASE-108. The LIT’s “Why this is here” says the exchange cited it for these. Neither entry now cites it. The paper gives a case of a choice settled by mutual inhibition between advocates and not by an arbiter. It says nothing about adversarial or compromised members, which CASE-108 is about.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Few swarms. The targeting result rests on two swarms with one island setup; the control on two more. Signals from the same bees are not independent, which the second test on minimum signaller counts only partly answers.
- The model is not in the paper. Its equations, parameters and the critical thresholds are in the supplement, and the main text gives them no values.
- How scouts tell rival dancers apart is unknown: dance decoding or odour, the authors say.
- Overstatement. The modelling results are called “conclusive”; whether the swarms observed were near the critical signalling rate is not shown.
- Errors. None of substance found; the three tests rebuilt from the text’s counts reproduce.
Open questions
- Is the observed rate of stop signalling above the model’s critical threshold, and by how much?