Try this first
Turn off the gossip.
Play the default seed. Set Sharing to Off and replay. The walkers keep their own short memory, but lose the neighboring traces. Does that change where they gather? Now try another seed before trusting the story.
What the knobs actually do
- Walkers / queries
- The swarm population, and the number of observations each policy receives per round. More walkers gives everyone more observations.
- Explore
- A chance to try another direction for a walker, or query a random location for the central learner. Same probability; different mobility.
- Sharing
- How often a walker follows the best remembered adjacent trace instead of its current heading. Traces fade 3% per round.
- Move food
- At round 81 the food shifts across the wrapping grid. No policy is told the new location.
- Seed
- A repeatable world, not a difficulty rating. Keep it fixed to compare one changed rule.
The crowd is not a tiny genius.
Its rules are: keep a direction that improved the last reward; sometimes follow a useful neighboring trace; explore again after repeated disappointments. The central learner instead remembers sampled rewards across the map and can return directly to the best known location.
That last advantage matters. This is a comparison of search policies and mobility—not proof that one brain is better than many. Both get the same number of observations; they do not get equal computation or equal movement constraints.
All the seeds, including the inconvenient ones.
Across seeds 1–32 at the defaults, mean reward per observation was 0.815 central / 0.506 swarm / 0.164 random. With sharing removed, the swarm fell to 0.226. Its trace-sharing rule helps here. It still does not beat the central learner.
We retain six complete batches—192 runs—including no relocation, high exploration, and different query budgets. No seed was dropped. Food never runs out, crowding costs nothing, and successful locations can be reused: a friendly world for centralized memory. Add depletion or costly movement and you have a new experiment, not permission to reuse this conclusion.
Original toy model. No insects, biological replication, language models, market data, investment advice or real-world performance claims.