A new computer model suggests that a shared directional code can emerge in bee-like colonies when finding food is neither too easy nor too difficult. The same simulations found that a direct-pointing code could shift gradually to a gravity-referenced code without communication breaking down across most of the tested conditions.
That is a result about computational feasibility, not a reconstruction of real bee evolution. The authors say the transition was not ruled out across nearly all tested evolutionary-parameter settings, but the model does not establish that the historical transition in Apis happened or estimate how likely it was.
When pointing became useful
The first stage asked when a direct-pointing referential code could take hold. Direct pointing evolved readily at an intermediate level of foraging difficulty, including settings with a few large food sites or many small ones. It failed when food was too sparse to seed dances or so abundant that random searching was already enough.
Each simulated population contained 60 colonies, with 80 workers in each colony. Each colony was evaluated over 50 independent episodes of 12 foraging attempts. The horizontal-comb stage ran for 60 generations across 50 replicate seeds, with the mutation scale fixed at , where the symbol denotes the size of simulated mutation steps. Parents were chosen in proportion to fitness, and copied traits received independent Gaussian mutation steps clipped to their allowed ranges.
The main measure was recruitment advantage: whether workers following a dance succeeded more often than contemporaneous random searchers. The researchers also tracked directional bias, foraging success and the tendency to dance in the final generation.
A difficult code switch
The second stage started all populations with a horizontal comb. It allowed comb tilt and orientation to evolve and activated sender and receiver transposition traits as part of a possible move to a gravity-referenced code.
The transition search ran for 120 generations. For each of two decoder designs, the researchers tested 1,024 parameter trials, using 10 random seeds per trial. A seed counted as stable when the final mean comb tilt reached at least 0.80 and both sender and receiver transposition means reached at least 0.50. A run was marked non-viable when its lowest foraging success was 0.02 or lower.
Stable transitions appeared with both decoders. In the optimization search, 61 trials using the flatten decoder were stable in all 10 seeds, compared with 33 using the unproject decoder. When the threshold was stability in at least eight seeds, the counts were 252 and 214, respectively.
The strongest clusters of stable trials combined abundant, large and reachable food with a strong incentive for vertical combs and a high mutation scale. The flatten decoder was concentrated in settings with larger patches and cheaper travel than the unproject decoder. In held-out tests, the strongest flatten candidate was stable in 95 of 100 seeds, while the strongest unproject candidate was stable in 90 of 100, and neither top candidate produced a non-viable seed.
The sensitivity analysis found different patterns for the two decoders. Mutation scale was the most sensitive parameter for unproject. For flatten, food-site count mattered most, followed by mutation scale, travel cost and maximum distance; the other tested parameters changed little.
What the model leaves out
The clearest broad evolutionary pattern was the modeled benefit of vertical combs. At fixed validated ecologies, stable-transition rates rose as that benefit increased for both decoders. Larger mutation steps and stronger correlation between sender and receiver changes were associated with higher rates within a given benefit level, but only partly offset weak benefit.
The model was deliberately narrow. It represented only the directional part of the dance, leaving out distance coding and differences in food quality. Colonies reproduced asexually, the vertical-comb benefit was an abstract fixed-form term rather than a mechanistic architectural model, and the mutation scale represented search over the simulated evolutionary period rather than a measured biological mutation rate. The model also blended transposition traits continuously, a choice that may enlarge the feasible transition region.
The document is an arXiv preprint, version 3, dated 12 September 2026. Its front matter says that the code, configurations and result files are available to reproduce every figure and table. The authors describe the findings as a feasibility result, not an estimate of natural likelihood or a reconstruction of actual Apis history.
Paper data and sources
Original title: The emergence and evolution of a referential code in populations of bee-like agents
Authors: Grzegorz Chrupała
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text