benchmarksopinionbullishDirectional representations shift the bottleneck from symbolic layer to neural layer, improving scaling potentialMachine Learning27 Jul 2026http://arxiv.org/abs/2607.02307v1