Harmony in Diversity: Multi-domain Contrastive Policy Optimization for Large Reasoning Models

Name: Harmony in Diversity: Multi-domain Contrastive Policy Optimization for Large Reasoning Models
Uploaded: 2026-10-03T23:16:43.449068+00:00
Duration: 404 s
Description: Multi-domain Contrastive Policy Optimization (MCPO) is a reinforcement learning method that improves large reasoning models by turning cross-domain interactions into knowledge transfer. It achieves this by structuring rollouts and using contrastive objectives to promote sharing of transferable knowledge across domains and consolidation of correct knowledge within each domain.

Multi-domain Contrastive Policy Optimization (MCPO) is a reinforcement learning method that improves large reasoning models by turning cross-domain interactions into knowledge transfer. It achieves this by structuring rollouts and using contrastive objectives to promote sharing of transferable knowledge across domains and consolidation of correct knowledge within each domain.

Open the full episode