Goodfire proposes a mechanistic shaping framework to use internal representations to suppress models from taking shortcuts
Shaping and Monitoring Models Through Their Internal Representations
AI BriefingGoodfire Research published a paper proposing the mechanistic shaping framework, which uses reliance shaping and expression shaping to penalize a model during training for relying on or expressing specific internal concepts, enabling it to solve problems correctly rather than merely arrive at the right answer.
Source: Goodfire Research · goodfire.com
Event follow-up1
View all event coverage- Goodfire proposes a mechanistic shaping framework to use internal representations to suppress models from taking shortcutsGoodfire Research · 11 days ago