kalomaze: Poor model transfer reflects narrow human problem selection, not failed generalization

kalomaze · x · 2026-09-12

kalomaze argues that 'models aren't transferring well' shouldn't be read as a model generalization failure, but as an artifact of how humans pick problems: we have bad intuitions about what helps broadly, so we scale up synthetic versions of what we already know to ask for—mostly the tasks SWEs care about—narrowing the training signal.

Related event: kalomaze: information asymmetry, not verifiability, drives RLVR generalization(9 posts)→

Original post →

More from Models

Models channel →