math-reasoning-transfer

Understand why math reasoning improvements don't always transfer to general capabilities. Use RL-based training instead of SFT to preserve representation structure and enable broader generalization.

Independent GitHub snapshot · Automated classification · Original documentation governs