Evaluators and safety organisations · Article · Goodfire
Shaping and Monitoring Models Through Their Internal Representations
Read at goodfire.com ↗Read at sourceNo licence statement found
Outline · 110 paragraphs
- Introduction¶8
- Mechanistic shaping through internal representations¶18
- Teaching a medical diagnostic model not to use confounders¶29
- Preventing LLM sycophancy where data filtering fails¶36
- Preventing reward hacking¶47
- Effects on model monitorability¶64
- An analytic theory of mechanistic shaping¶72
- A five-level hierarchy for monitorability¶79
- Empirical observations¶84
- Looking forward¶94
- Citation¶99
- Footnotes¶101