CADENCE ARCADE

Cadence (cadence-net, by Pragma Research) is a brain-like learning library: regions of neurons that settle on every answer and learn with a local rule, no backprop. These are independent experiments that put it in games and report what actually happened, misses included.

Mario jumping up the stairs at the end of World 1-2, played by the Cadence brain W1b
Super Mario Bros. · main showcase

One brain beat all of World 1 in a single 5-life gauntlet

Brain W1b cleared 1-1, 1-2, 1-3 and 1-4 (Bowser) back to back with the same weights, alone from the screen. Not consistent yet: 7 to 27 clears per level on 50 held-out starts.

Clips, brains and scoreboard →
Brain-cam frame: Mario in 1-4 next to the brain's recorded neuron activity
Mario · interactive

Brain scope

Scrub through solo runs and see the brain's real recorded neuron activity at every decision, or poke it to see what each move depended on.

Open the scope →
Mario underwater in World 2-2, the blind brain's first sight of the level
Mario · first sight

Mario, level by level

One brain plays the levels in order, learning as it goes. At first sight of a new level it never beat the best simple baseline.

Results →
The chat page: a short exchange with the Cadence brain
Toy language

Chat

Text one Cadence brain on a toy 188-word grammar. 83% valid replies on held-out prompts; a plain MLP gets 95%.

Chat with it →
Ember Field replay: two lantern creatures under falling embers
Original game

Ember Field

Learn which embers feed and which sting, then the colours swap. The Cadence brain relearns after each turn; a plain neural net is the stronger player here.

Watch the replay →
One-Frame Breakout replay boards
MinAtar · cadence-net 0.89.0

One-Frame Breakout

One frame at a time, so the ball's direction is hidden. The new recall position term didn't help as shipped (6 seeds better, 6 worse).

Replay and verdict →
Chart of bricks per game over a 10x longer training run
MinAtar · follow-up

Breakout, 10x longer

The same Breakout players trained 10x longer. Cadence keeps climbing; the two-frame MLP pulls ahead.

Results →

Older Mario pages: the first 1-1 run (10/28-29, replay and learning curve) and watch it as plain video.