Adaptation-Mode Deep Dive

Back to results

This deep dive records paired adaptation-mode comparisons across the completed Apr24 grid.

For matched strategy router score, lora - full has mean delta -0.0116 and median delta +0.0021 over 78 matched pairs. reinit - full has mean delta -0.0967 and median delta -0.0722. reinit - lora has mean delta -0.0852 and median delta -0.0410.

For final answer accuracy, lora - full has mean delta +0.0069 and median delta +0.0036. reinit - full has mean delta -0.0823 and median delta -0.0171. reinit - lora has mean delta -0.0892 and median delta -0.0262.

For router probe alone, the paired deltas are smaller and not aligned with the matched-score pattern: lora - full has mean delta -0.0200, reinit - full has mean delta +0.0029, and reinit - lora has mean delta +0.0228.

For baseline-normalized reconstruction, the mean deltas are small: lora - full is +0.0059, reinit - full is +0.0130, and reinit - lora is +0.0071.

Open empirical checks:

• Split the adaptation deltas by raw versus normalized loss settings, since raw runs dominate the bottom individual matched-score list.
• Inspect whether reinit loses matched score mostly through strategy match rate, final answer accuracy, or interaction between the two.
• Compare adaptation modes under matched initial reconstruction loss or matched early validation accuracy.