Curriculum-Guided Automated Research: Progressive Complexity for LLM Ideators

TL;DR: Instead of dropping LLMs straight into full-blown research, what if we train them on a curriculum that gradually ramps up idea complexity, using execution feedback to shape the path? This could help avoid early saturation and foster skill transfer.

Research Question: Does curriculum learning—where the complexity of research tasks/idea spaces is gradually increased based on execution outcomes—lead to better long-term ideator performance and less saturation in automated AI research?

Hypothesis: A curriculum-guided ideator will develop more generalizable research heuristics, avoid early plateaus, and maintain innovation as the environment complexity increases, compared to direct exposure to the hardest tasks.

Experiment Plan: Design a curriculum of research ideation tasks, starting from toy/low-dimensional problems and scaling up to full LLM pre-training/post-training. Use execution feedback to adapt the curriculum pace (move forward when performance saturates). Measure ideator’s transfer learning, saturation points, and solution diversity against non-curriculum baselines.

References:

• Chen, J., Liu, F., Liu, N., Luo, Y., Qin, E., Zheng, H., Dong, T., Zhu, H., Meng, Y., & Wang, X. (2025). Step-wise Adaptive Integration of Supervised Fine-tuning and Reinforcement Learning for Task-Specific LLMs. arXiv.org.
• Si, C., Yang, Z., Choi, Y., Candès, E. J., Yang, D., & Hashimoto, T. (2026). Towards Execution-Grounded Automated AI Research.