Search papers, labs, and topics across Lattice.
Cohere Labs
1
0
2
Confidence calibration in reasoning models can be dramatically improved by aligning supervision with the model's state, cutting ECE by over half on challenging benchmarks.
