This paper introduces a higher-order attention layer operating at zero temperature, where state resides on item pairs (a, b) and third items k serve as witnesses evaluated via legs (a, k) and (k, b). The mechanism retains only the best witness and its tied count, propagates bits comparing these counts, and reconstructs pair state from input and bits each loop, trained solely from answers. This design eliminates three soft-attention failure modes at scale: dispersed weight, unresolved score gaps, and drifting soft state. In maximum retrieval, the zero-temperature layer achieves ≥99.95% accuracy on sets 1,000× larger than training across all ten seeds, whereas a soft-count variant drops to 42% at 4,096 items. On CLRS-30 without hints, a general form (half channels at zero temperature, embedded numbers, carried state) surpasses the best published hint-using model; the full cold layer is exact on all test inputs for 9 of 12 tasks where answers depend on numbers only through comparisons. From yes/no supervision, it solves 2-SAT and two-coloring exactly at 500 variables, but not parity equations, matching theoretical predictions. Adjacent to a frozen 7B language model, it chains family relations exactly at 1,000 statements after training on 16, where the LM alone falls to 44%; finite weight checks prove every block that learned correctly generalizes to all sizes on acyclic contexts given the LM's reads. On CLUTRR with given graphs, it achieves 100% accuracy on 5–10 relation stories, learning all 62 rules exactly; from text with the LM frozen and only a small read-in trained, it lifts an open 7B model from 33% to 65%, exceeding all non-frontier systems without hand-written rules.

Read original