On-Policy Delta Distillation
On-policy distillation is an alternative post-training method in reinforcement learning that alleviates the constraints imposed by reward models by providing token-level supervision from a teacher model. Although on-policy distillation has been studied and applied across various settings, its fundamental design remains underexplored. In this paper, we introduce a new distillati
Reading in Plain English mode — toggle below to switch
On-policy distillation is an alternative post-training method in reinforcement learning that alleviates the constraints imposed by reward models by providing token-level supervision from a teacher model. Although on-policy distillation has been studied and applied across various settings, its fundamental design remains underexplored. In this paper, we introduce a new distillati
This model release advances Reasoning and Benchmark, directly affecting how engineers build, optimize, and ship AI systems. Teams running production LLM workloads need to understand these techniques to stay competitive in hiring and architecture decisions.