Meta-DynaAct: Meta-Learned Dynamic Action Space Construction Across Diverse Domains

Research Question: Can meta-learning be leveraged to create a generalized policy for dynamic action space construction that adapts efficiently to new reasoning domains and problem types?

Hypothesis: A meta-learned DynaAct variant will outperform standard DynaAct and static baselines in out-of-domain generalization, constructing more effective action spaces for previously unseen tasks with minimal tuning.

Experiment Plan: - Collect a diverse set of reasoning benchmarks (e.g., math, code, planning, chemistry).

• Implement a meta-learning framework (e.g., MAML or RL^2) where the "learner" receives task features and past performance metrics, and outputs dynamic action space construction strategies (parameters for submodular functions, diversity/utility trade-offs, etc.).
• During meta-training, expose the learner to many tasks; during meta-testing, evaluate on held-out tasks.
• Compare action space quality (measured by downstream task accuracy and inference efficiency) to DynaAct and static methods.
• Analyze transfer and adaptation speed.

References: ['Zhao, X., Wu, W., Guan, J., Li, Q., & Kong, L. (2025). DynaAct: Large Language Model Reasoning with Dynamic Action Spaces.', 'White, C., Safari, M., Sukthanker, R., Ru, B., Elsken, T., Zela, A., Dey, D., & Hutter, F. (2023). Neural Architecture Search: Insights from 1000 Papers. arXiv.org.']