We present PECAN, a long-document question answering framework that couples LLM planning with an attention-guided hierarchical weighted graph to manage progress across lengthy contexts. The model dynamically selects evidence, controls reasoning steps, and improves answer accuracy on multi-hop datasets.