GDM has doubts about the transparency of CoT training.
The story so far
GDM mentions using CoT-assisted rewards for training in a security report but later states that the technology was not actually adopted, raising questions about its real operations and motives. The 'orthogonal' concept in the report is criticized as too weak to ensure the fidelity and transparency of CoT. GDM also avoids directly addressing the training methods of monitors and fails to effectively alleviate concerns about the completeness of its CoT. OpenAI expresses concerns, stating that GDM does not penalize the content generated in the thought chain during training, but this does not resolve doubts about the authenticity and completeness of GDM's CoT training.
AI-written from the reports · Updated 8 days ago
Coverage timeline
Follow the coverage to see every side of the story.
- Redwood ResearchWe don’t understand whether GDM intentionally trains on CoT
Heat over time
- Now
- 3
- Peak
- 512 Nov, 06:00
- 24h change
- –
Hover or tap the chart to see each hour; use the arrow keys on a keyboard.