A computer model of evolving bee-like colonies found that direct pointing appeared only in a middle band of simulated foraging conditions. It was not favored when food was so sparse that dances could not be seeded, or when food was abundant enough for random search to suffice.
The same simulations found stable transitions from the direct code to a code using gravity as its reference for direction, but only under selected combinations of settings. The strongest candidates were then tested on fresh random seeds in held-out validation.
The first signal had a narrow ecological window
On a flat comb, recruitment advantage rose above baseline only in a diagonal band of food ecologies. Direct pointing did not evolve below a patch radius of a few tens of meters. The advantage was greatest with a few large patches, then declined for small patches and for abundant large patches.
Evolution happened at the colony level
Evolution took place at colony level, while behavior came from individual bee-like agents. Colonies reproduced asexually and passed on heritable traits subject to mutation. Each simulated population contained 60 colonies with 80 workers each, and every colony was evaluated over 50 independent foraging episodes with 12 attempts per episode.
In the direct-pointing stage, comb tilt was fixed at zero. Colonies evolved for 60 generations with a mutation scale of 0.07, while food distribution was varied across 50 replicate seeds for each condition.
For the transition search, the model varied eight parameters and compared two decoders, called flatten and unproject. It ran 1,024 optimization trials per decoder. The first 64 trials were sampled at random, and the search used a Tree-structured Parzen Estimator implemented in Optuna.
The study counted a transition as stable only if final mean comb tilt reached at least 0.80 and both sender and receiver transposition means reached at least 0.50. A run was non-viable if its lowest foraging success was 0.02 or lower.
The transition had to survive repeated tests
Across the 1,024 trials for each decoder, 14 flatten trials and 24 unproject trials were stable in all 10 seeds. At least one stable seed appeared in 611 flatten trials and 641 unproject trials.
The top five candidates from each search were rerun on a 40-seed confirmation panel and a 100-seed held-out validation panel. Stable rates ranged from 80 to 91 of 100 for flatten and 82 to 92 of 100 for unproject, with no non-viable seeds reported.
The strongest validated candidate for each decoder achieved stable transitions in 91 of 100 flatten seeds and 92 of 100 unproject seeds. Because the study is computational, these percentages describe the selected simulations rather than measured rates in a natural bee population.
Mutation scale was a sharp boundary
Mutation scale marked the sharpest one-parameter boundary in the sensitivity tests. Under flatten, stable transitions fell from 91 of 100 seeds to 49 of 100 when the mutation scale was set to 0.05. Under unproject, the rate fell to 16 of 100 at 0.04.
With validated ecology held fixed, stable-transition rates were robust near a modeled vertical-comb benefit of about 0.56 to 0.60, marginal at 0.30 and collapsed at 0.10. Higher mutation and sender-receiver correlation only partly compensated for the weaker benefit.
At the smallest mutation scale, the stable rate at sender-receiver correlation 0.9 was roughly twice the rate at correlation 0.
A feasibility result, not a history of bees
Taken together, the simulations describe a conditional feasibility result. Direct pointing appeared in a bounded ecological window, and stable transitions were found under some combinations of model settings. The work does not show that the transition happened in actual bee evolution or estimate its historical likelihood.
The reported rates belong to the model’s selected conditions. The direct-pointing stage fixed comb tilt at zero, and the transition search relied on two decoder choices and a selected parameter search. They should therefore be read as simulation outcomes under those settings, not as a historical rate for bee evolution.
The work is an arXiv preprint, version 1, dated 26 Aug 2026. The document says code, configurations and result files for reproducing every figure and table are available at https://github.com/gchrupala/bees. No separate supplement or funding statement is reported.
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