Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation

Key points

• Deep-search agents are commonly trained by distilling tool-use trajectories from stronger teacher models.1
• However, structurally complex questions may remain relatively easy for teachers familiar with key entities in the underlying entity graph, resulting in lower search demand and weaker supervision.1
• We refer to this teacher-specific variation in question difficulty as teacher-relative difficulty and investigate entity familiarity as an early predictor.1
• To address this limitation, we propose Self-Reported Entity Familiarity (SEF), a lightweight gating signal that directly asks a teacher whether it is familiar with an entity before admitting that entity into the graph.1

Sources (1)

• [1]Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 06:25 AM

  We first introduce Evidence-Verified Entity Familiarity (EVF), which estimates familiarity by verifying a teacher-generated factual brief against online evidence.

  Deep-search agents are commonly trained by distilling tool-use trajectories from stronger teacher models.

Extractive summary: sentences quoted from the sources.