- What changed
- Cohere introduced CALIBER, a training method for reasoning models that assesses success probability before and after reasoning, reducing calibration error on math benchmarks.
- Why you should care
- Standard reasoning models often fail to reliably communicate uncertainty despite generating detailed paths.
- Your move
- Watch. Monitor how pre and post reasoning confidence calibration is adopted across other model architectures.
- What to watch next
- Publication of peer reviews or open source code release for CALIBER.
- Event
- research
- Event date
- Oct 20, 2026
- Relevant to
- General AI readers
