A new preprint on arXiv takes a critical look at self-distillation in agentic reinforcement learning. The authors note that the field has produced a growing number of methods, but the way these methods are evaluated may be undermining the ability to judge their true value.
The paper's abstract points to a specific problem: existing evaluations often reuse published baseline scores that were generated under different experimental conditions. This practice, combined with a lack of direct comparisons between methods, makes it difficult to determine whether reported improvements are real or artifacts of the setup.
The abstract is truncated, so the full details of the proposed remedy are not available from the source. Still, the core message is clear: the current evaluation culture in this niche of RL research needs closer scrutiny before the community can trust the effectiveness of new self-distillation techniques.