Analyzer instruction generation without reference trajectories

Determine whether a substantially stronger analyzer could write effective Privileged Self-Practice instructions from student failures alone, without access to a reference trajectory.

Background

In Privileged Self-Practice, the analyzer normally receives both the student's failed rollouts and a reference trajectory. The reference trajectory supplies information about a successful procedure and the likely reason for failure, while the student rollouts provide task-specific evidence about the current policy's mistakes.

An ablation removes the reference trajectory and shows a performance decline, especially for Qwen3-4B. The paper therefore leaves unresolved whether a much larger analyzer could infer sufficiently effective task instructions from failures alone, without the reference trajectory.

References

Whether such an analyzer could write effective instructions from the failures alone without a reference trajectory, is left to future work.

— From Self-Distillation to Self-Practice: Privileged Information for Multi-Turn Agents  (2609.29051 - Su et al., 24 Sep 2026) in Appendix, Section Experimental Details, subsection Reference trajectories, paragraph “Why Construct Reference Trajectories.”