A new training method fixes a specific way that shrinking AI agents down to size backfires.
Researchers describe a technique called Guided On-Policy Distillation, or Guided-OPD, for training smaller "student" models to copy larger "teacher" models on multi-turn tasks - the kind where an AI plans steps, calls tools, and reacts to an environment over several turns. The standard method for this, on-policy distillation, breaks down in that setting: small student mistakes compound turn after turn, pulling the trajectory into territory the teacher model has never seen, so the teacher's guidance gets worse exactly when the student needs it most. Guided-OPD fixes this by mixing teacher-generated and student-generated turns within each training rollout, then gradually reducing how often the teacher steps in until training matches how the model actually runs at inference. Tested on three simulated-environment benchmarks - ALFWorld, ScienceWorld, and WebShop - using Qwen3 students distilled from a larger Qwen3-30B-A3B teacher, the method beat standard distillation by 21.1 percent on task score and 25.5 percent on success rate on average, with the smallest student models improving the most.
Most companies can't afford to run the biggest models for every step of an AI agent's work, so distilling that capability into smaller, cheaper models is what makes agentic AI viable at scale. This paper isolates a specific failure mode - errors compounding across turns and pushing the student outside the teacher's comfort zone - rather than just tuning data volume or teacher size, and the fix pays off most for the smallest, cheapest models, which is exactly where the cost pressure is highest.
The gains are measured against three text-based simulation benchmarks, not messy real-world deployments, so whether this curriculum trick holds up when tasks get open-ended and the teacher's own judgment gets shakier is still an open question.