Cross-Pollination of SFT Lessons Boosts AI Alignment and Generalization

A new paper by Anton de la Fuente and Arthur Conmy argues that supervised fine-tuning (SFT) techniques developed in one AI subfield—alignment training, model organisms, or toy models—can often be transferred to others, accelerating progress across the board. The authors demonstrate three concrete transfers. First, they port a lesson from alignment into toy models: training on the reasoning behind a behavior (as in Anthropic's 'Teaching Claude Why') leads to broader generalization than training on examples alone. Second, they port a capability preservation lesson from model organisms into the Model-Spec Midtraining alignment setting: using outputs from a model other than the student (off-model) for SFT damages capabilities, but mixing in benign on-model (on-policy) data prevents most of that damage while still embedding the target behavior.

Third, the researchers port a robustness lesson from model organisms into the same alignment setting: following up alignment SFT with benign SFT can erase the alignment behavior while preserving capabilities, proving that capability preservation alone does not ensure robustness to subsequent training. This result highlights a critical vulnerability in current fine-tuning pipelines. The paper emphasizes that many SFT projects share underlying goals, and insights from one area—such as how to prevent capability degradation or how to induce robust behavioral changes—should be systematically tested across others. By demonstrating concrete cross-domain porting, the authors encourage AI researchers to borrow techniques from outside their immediate specialty, potentially uncovering universal principles that make alignment more efficient and reliable.