No training
Taste and feed
Sugar drops drunk; bitter drops refused
| Real | Rewired | Random | |
|---|---|---|---|
| Sugar drops drunk | 28 | 0 | 22 |
| Bitter drops drunk | 0 | 0 | 11 |
| Sugar with bitter drunk | 0 | 0 | 3 |
Home / Benchmark
Benchmark: one fly circuit, eight tasks, four brains
Feeding, backing away from walls, poker, Pong, an odour maze, a two flower gamble, gusty flight and a chase. Every task runs on the same 1,846 neurons cut from the FlyWire connectome, and every result is shown next to the same task on rewired, random and silenced circuits. The question is never only whether the fly can do it, but whether the real wiring is what makes it possible.
How a task works
Senses drive real sensory neurons: taste neurons for sugar and bitter, olfactory receptor neurons for odours, visual projection neurons for objects and looming, HS cells for self motion. Activity then spreads through every measured synapse between them and the descending neurons that command the body, plus MN9, the proboscis motor neuron.
The two behaviours need no training: the wiring alone has to produce them. The six other tasks train a small linear readout of 21 output neurons with the cross entropy method; the circuit never changes. The benchmark trains the real circuit and each control with the same seed, population and number of generations, and scores them on settings they never trained on. So far the real wiring beats every control on 5 of 6 trained tasks.
Real fly behaviours
Reflexes of the real circuit, with no training at all: the wiring alone has to produce the behaviour.
Games
Games a fly never evolved for. A trained readout turns the circuit's activity into moves.
Trained readout
Kuhn poker
Expected chips per hand against an opponent it never met (0.30 is the best possible)
| Real wiring | 0.01 | |
|---|---|---|
| Rewired, same degrees | -0.02 | |
| Random wiring | -0.02 | |
| Circuit silenced | -0.02 |
Against the equilibrium player: 0.000 chips per hand (0 is the best possible).
Real wiring beats every control
Trained readout
Pong
Hits minus misses with a faster ball
| Real wiring | 2.88 | |
|---|---|---|
| Rewired, same degrees | -2.31 | |
| Random wiring | -1.43 | |
| Circuit silenced | -3.69 |
Real wiring beats every control
Cognitive tests
Choice and memory. Where a circuit without learning reaches its limits, the results say so.
Trained readout
Odour T-maze with a delay
Choice confidence towards the right arm (1 perfect, 0 chance)
| Real wiring | 0.03 | |
|---|---|---|
| Rewired, same degrees | 0.04 | |
| Random wiring | 0.03 | |
| Circuit silenced | 0.00 |
No circuit gets above chance
Trained readout
Two flowers
Expected reward per visit (0.5 chance, 0.8 best)
| Real wiring | 0.57 | |
|---|---|---|
| Rewired, same degrees | 0.50 | |
| Random wiring | 0.50 | |
| Circuit silenced | 0.50 |
Real wiring beats every control
Body and movement
Steering in flight and on foot, with the realistic fly body.
Trained readout
Hold a course in flight
Mean cosine of the heading error in strong gusts (1 on course)
| Real wiring | 0.97 | |
|---|---|---|
| Rewired, same degrees | 0.63 | |
| Random wiring | 0.09 | |
| Circuit silenced | 0.26 |
Real wiring beats every control
Trained readout
Chase a moving target
Mean closeness to a faster target (1 touching)
| Real wiring | 0.57 | |
|---|---|---|
| Rewired, same degrees | 0.37 | |
| Random wiring | 0.15 | |
| Circuit silenced | 0.14 |
Real wiring beats every control
Honest limits
Sensory encodings for games are arbitrary: the fly does not know what a card is; the jack, queen and king simply drive three different groups of visual neurons. The readout is linear, so any skill beyond a lookup has to come from the circuit. There is no learning inside the circuit, so tasks that need memory across seconds are expected to sit near chance, and the page says so when they do. Every number comes from web/scripts/gym-benchmark.ts and can be reproduced.