What Should SFT Learn? Conditional Information Gain for Generalizable Post-Training

Supervised fine-tuning (SFT) remains a widely used approach for post-training large language models, because fixed demonstrations can directly introduce desired reasoning strategies and behavioral patterns.