Composition Collapse: Stable Factual Knowledge Does Not Imply Compositional Reasoning

In three linesarXiv paper reveals that models with statistically indistinguishable atomic knowledge fail systematically to chain them in multi-hop reasoning (>40 percentage point gap). Aggregate metrics mask this 'composition collapse'. Authors introduce a double-gate protocol decomposing post-training gains into three independent channels: atomic stability, residual composition, and critical depth.

Summary generated by Claude — human-verified