HIPIF: Hierarchical Planning and Information Folding for Long-Horizon LLM Agent Learning

The article introduces Hierarchical Planning and Information Folding (HIPIF), a novel approach for enhancing long-horizon learning in Large Language Models (LLMs) by addressing long-context interference through subgoal decomposition and progress summarization. HIPIF employs end-to-end training that organizes tasks around explicit subgoals while folding completed histories to mitigate interference, and incorporates hierarchical reflection and subgoal-oriented process rewards for stability. This method shows promise in improving performance on multi-turn tasks, which is crucial for practitioners aiming to build more effective autonomous agents in complex environments.