TL;DR of the paper, generated automatically.

Tweaking a few parameters boosted some benchmark scores, but results were uneven.

• Local edits: 1,023 parameters changed; up to 29.45-point MATH-500 gain.
• Transfer: a math-edited checkpoint improved GPQA, IFEval, SciBench, not SciCode.
• Reliability: six of 16 trajectories kept original weights when fixes couldn't be validated.