The paper introduces a technique called information-gain calibrated on-policy distillation, aimed at improving how large search agents are compressed into smaller, more efficient student models. On-policy distillation has already shown promise for this transfer, but the new work focuses on deep search settings, where agents must repeatedly pull in external evidence to guide their next steps.

According to the abstract, the method specifically calibrates the distillation process using information gain. This likely means the student model is trained to prioritize evidence that most reduces uncertainty, rather than simply mimicking the teacher's actions. However, the abstract is truncated, so the exact mechanism and experimental results are not fully described.

Overall, the work points toward more sample-efficient and targeted distillation for complex, iterative reasoning tasks. It remains to be seen how the calibration affects performance compared to standard on-policy distillation, as those details are not available in the provided source.