SafetyPage

Evaluators and safety organisations · Blog post · Redwood Research

Towards understanding score-seeking in LLMs

Read at blog.redwoodresearch.org ↗Read at sourceAll rights reserved

Outline · 59 paragraphs

  1. Methodology¶13
  2. Key results¶16
  3. Tested models seek score terminally¶18
  4. Models violate instructions to get a high score¶19
  5. Models care about score regardless of downstream consequences¶22
  6. Do models help other agents achieve a higher score?¶28
  7. What this means for alignment risk¶38