Meta-Self-Evolution: Learning to Schedule Guidance vs. Autonomy for Task Generalization

TL;DR: Train a meta-controller that learns when to apply human-grounded guidance versus fully autonomous self-evolution, aiming for optimal performance and lowest supervision across diverse domains. The experiment would involve meta-training R-Few across a battery of tasks, letting the controller adjust the guidance schedule to generalize well even to novel tasks.

Research Question: Can a learned meta-controller optimize the tradeoff between guidance and autonomy in self-evolving LLMs, improving generalization and minimizing supervision?

Hypothesis: A meta-learned scheduler that dynamically balances guidance vs. autonomy outperforms fixed or heuristic schedules, achieving strong generalization with minimal human data.

Experiment Plan: - Train a meta-controller (e.g., via reinforcement learning or Bayesian optimization) to adjust guidance frequency in R-Few based on model progress, task type, and drift signals.

• Evaluate on a suite of reasoning, math, and domain-specific tasks (some unseen during meta-training).
• Compare generalization, label efficiency, and stability to fixed-schedule R-Few and fully autonomous approaches (e.g., Crescent, Sun et al., 2025).
• Expected outcome: Meta-scheduling achieves the best tradeoff between autonomy and stable self-evolution.

References:

• 1. Sun, Y., Chen, M., Zhao, T., Xu, R., Zhang, Z., & Yin, J. (2025). The Self-Improvement Paradox: Can Language Models Bootstrap Reasoning Capabilities without External Scaffolding? Annual Meeting of the Association for Computational Linguistics.
• 2. [R-Few paper cited as source]