AgentBrew: Lifelong Knowledge Brewing from Strong Teachers to Weak LLM Agents
Yangqin Jiang, Chao Huang
Why It Matters
What makes this one worth your time
This approach could significantly enhance the performance of LLM agents in practical applications by leveraging existing expertise without the need for extensive retraining.
AgentBrew enables effective knowledge transfer from strong teachers to weaker LLM agents without retraining.
Summary
The paper introduces AgentBrew, a method for distilling knowledge from a strong teacher to a weaker LLM agent without requiring weight updates or expert demonstrations, addressing challenges of sparse feedback and tailored guidance.
Key contributions
- Introduction of the AgentBrew framework for knowledge brewing from teachers to students.
- Development of the failure-triggered teacher-Ralph Loop for generating actionable notes from student failures.
- Implementation of student-aware synthesis for calibrating teacher knowledge to weaker agents.
Notable insights
- The failure-triggered teacher-Ralph Loop effectively transforms student failures into actionable guidance, which is a novel approach to dealing with sparse feedback.
- Student-aware synthesis ensures that the distilled knowledge is tailored to the operational capabilities of weaker agents, enhancing practical applicability.
Possible limitations
- Not stated in the abstract.
Abstract
arXiv:2607.16851v1 Announce Type: new Abstract: Deploying LLM agents typically requires a compact test-time student, even if a stronger teacher is available during training. We study knowledge brewing: distilling a teacher's interactive experience into a persistent external memory for the student. Crucially, this requires no weight updates, expert demonstrations, ground-truth labels, or test-time teacher access. This setting poses two challenges: environments provide only sparse, binary feedback, and teacher-authored notes must be inherently tailored to be concretely executable by a substantially weaker student. To address these hurdles, we propose AgentBrew, comprising two coupled components. First, a failure-triggered teacher--Ralph Loop mitigates sparse feedback by transforming student failures into environment-validated notes. Second, student-aware synthesis calibrates teacher knowledge to the weak executor's operational granularity, yielding model-specific, actionable guidance. Extensive evaluations and comprehensive ablations across coding, math, and tool-use tasks demonstrate that this asymmetric, training-free brewing paradigm produces highly capable yet deployable LLM agents.