Attention at absolute zero

Cold attention: a looped block of about 150,000 parameters, trained on inputs of at most 16 items, running here in your browser on all 30 algorithms of CLRS-30, trained without hints. One general block shared by all 30 tasks averages 97.20% at 64 items over three seeds, against 96.64% for FloydNet (one model per task); one general block per task, 97.04%; each task in the setting the theory prescribes, 97.98% on the benchmark's standard inputs.

Bellman-Ford

Each node points to its predecessor on a shortest path from the source.

General block: the input's numbers enter as embeddings and the state is carried from loop to loop; read after 8 loops.

1 / 0
chosen by the general block, rightly chosen, wrongly missed: in the true answer
loops (read after 8)

Dots: share of the answer right when read after that loop.

thirty tasks, test at 64 items

Rose: the general block, one per task, the same configuration on all 30 (the bar, from 60 to 100). Cobalt: the cold block where the task allows it (numbers compared, pair numbers, or read after n loops; on strongly connected components it is below the general block). Right, and the tick: FloydNet, 20 to 215 million parameters per task.