Relative Representations Enable Portable White-Box Monitoring Across Language Models

Key points

• We show that models do in fact organize information similarly by using specific behavioral probes that operate on compatible representations.1
• We introduce a global anchored construction that summarizes information from all transformer layers and removes model-specific layer selection, so the same frozen probe can be straightforwardly applied to another model.1
• We evaluate cross-model probe transfer across monitoring tasks on trivia datasets, safety monitoring, and mathematical reasoning.1
• Overall, global anchoring reveals substantial task-relevant compatibility across independently trained models and provides a practical basis for portable white-box monitoring.1

Sources (1)

• [1]Relative Representations Enable Portable White-Box Monitoring Across Language Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 03:26 PM

  Independently trained autoregressive language models often acquire similar capabilities, but it remains unclear whether their internal representations organize behaviorally relevant information in compatible ways.

  We show that models do in fact organize information similarly by using specific behavioral probes that operate on compatible representations.

Extractive summary: sentences quoted from the sources.