IDRL
A post-training paradigm that alternates between online policy distillation and reinforcement learning within a single training loop. It aims to simultaneously improve deep reasoning ability and long-horizon agent task performance while enhancing knowledge transfer efficiency, optimization stability, and credit assignment.
Timeline (last 90 days)
Pistis's training pipeline is divided into two stages: the first stage is large-scale multimodal supervised fine-tuning (SFT), and the second stage is Interleaved Distillation and Reinforcement Learning (IDRL).
IDRL completes optimization within a single training loop by alternating between distillation and reinforcement learning objectives.
IDRL's alternating mechanism yields three benefits: more effective knowledge transfer, stronger optimization stability, and more precise credit assignment for long-horizon agent trajectories.
All Facts (3)
Pistis's training pipeline is divided into two stages: the first stage is large-scale multimodal supervised fine-tuning (SFT), and the second stage is Interleaved Distillation and Reinforcement Learning (IDRL).
50%UnverifiedIDRL completes optimization within a single training loop by alternating between distillation and reinforcement learning objectives.
50%UnverifiedIDRL's alternating mechanism yields three benefits: more effective knowledge transfer, stronger optimization stability, and more precise credit assignment for long-horizon agent trajectories.
50%