The networks on this bench.

What follows is only what makes each one unusual, not how it works. For that, read the whitepaper and the source.

Reverire, the hypervector network.

The learned state is one or more dense real vector fields, shaped by ordinary gradient descent. A layer holds a few vectors of the state width and moves the state, rather than mapping it through a weight matrix into a new space.

Each layer starts as the identity and has to earn its correction, so depth is resolution available to the task instead of computation the task must pay for at every step. Mixing inside a layer is a learned reflection, so a layer costs the state width rather than its square.

For sequences, order is carried by repeated geometric transport of one state. There is no positional encoding and no attention matrix, and the state stays the same size at any length.

It is not published, and this bench is the first evaluation of it against other architectures under a fixed budget.

Whitepaper and source

GAHP, the exclusion network.

It stores no weights. The only persistent state is a list of what has been ruled out, and fixed disorder proposes a form from whatever survives. The gradient decides what to remove, never where to move.

Nothing is ever reinforced. A move or a label leaves a table once the lower bound of its blame rate there is confidently above a threshold, so competence appears as the bad possibilities disappear.

Bans are not permanent. A revision pass gives every ban a reputation, tests the weakest one against what the table proposes now, and lifts it when the ban is doing worse than the survivor.

Whitepaper and source

Reading further.

The whitepapers carry the equations and the derivations; the repositories carry the implementation, the tests and the experiment logs behind every number on this bench. Both are worth reading before drawing conclusions from the tables.

Results so far: Go, 5x5, from self play