Teaching the Harness: Student-Aware Skill Transfer for LLM Agents
Summary
- Researchers introduce Student-Aware Teaching (SAT), a framework where a stronger LLM teacher builds a learner model of a student agent to transfer skills effectively.
- SAT uses diagnostic probes and empirical profiles to test hypotheses about student limitations and writes reusable skills for diagnosed failures.
- In experiments with Claude-Sonnet-5 teaching DeepSeek-V4-Flash, SAT achieved 86.46% task success, outperforming profile-free teaching (81.04%) and the untaught student (75.00%).
- Saved skill libraries maintained average performance gains of up to 11.1 points after the teacher departed.
Why it matters
This research demonstrates how tailoring skill transfer to specific LLM agent limitations and resource constraints significantly improves multi-agent training efficiency and task success rates.