Curriculum-Folding: Dynamic Curriculum Creation via Context-Folding and Difficulty Estimation for Ultra-Long-Horizon LLM Agents

TL;DR: Imagine if an agent could not only split long tasks into smaller parts but also decide which “chunks” to learn first, gradually scaling up difficulty—like starting with easy puzzles and leveling up as it gets smarter. The idea is to blend KLong’s trajectory-splitting with adaptive curriculum learning, where a context-folding mechanism actively manages and schedules training on easier-to-harder sub-tasks based on agent performance and task complexity.

Research Question: Can combining adaptive curriculum learning with context-folding and trajectory-splitting yield more efficient and robust training for LLM agents tackling ultra-long-horizon tasks?

Hypothesis: Integrating an adaptive curriculum—where the agent first trains on easier, context-folded sub-trajectories and progressively advances to harder, more complex ones—will accelerate convergence and improve generalization, compared to uniform or static progressive RL schedules.

Experiment Plan: - Set up an LLM agent using the KLong framework, but replace the static progressive RL schedule with a dynamic curriculum manager.

• Implement context-folding (as in Sun et al., 2025) to summarize and collapse completed sub-tasks, feeding difficulty estimates into a curriculum controller.
• Use metrics like sub-task success rate and learning progress to adaptively select the next set of training sub-trajectories.
• Measure convergence speed, success rates on PaperBench and MLE-bench, and generalization to unseen long-horizon tasks.
• Compare to KLong’s static progressive RL and curriculum-free baselines.

References:

• Sun, W., Lu, M., Ling, Z., Liu, K., Yao, X., Yang, Y., & Chen, J. (2025). Scaling Long-Horizon LLM Agent via Context-Folding. arXiv.org.
• Wu, Y., Zhang, J., Hu, N., Tang, L., Qi, G., Shao, J., & Song, W. (2024). MLDT: Multi-Level Decomposition for Complex Long-Horizon Robotic Task Planning with Open-Source Large Language Model. International Conference on Database Systems for Advanced Applications.
• Kar, I., & Kumar, K. C. K. (2025). Curriculum Guided Massive Multi Agent System Solving For Robust Long Horizon Tasks. arXiv.org.