Failure Modes in Parametric Graph Retrieval for Multi-Hop Question Answering

Graph-based retrieval improves multi-hop question answering by exposing evidence paths across documents, but explicit graph traversal can intro- duce additional inference-time cost and failure modes. We study whether the routing behavior of a BFS-based graph retriever can be approximated by a compact MLP reranker. On a 400-question HotpotQA distractor dev set, the MLP-pruned system remains competitive with full BFS, slightly improving F1, supporting-fact recall, bridge-node recovery, and latency, while giving up a small amount of exact match. We also report a 100- question pilot ablation used to study candidate caps, latency, hub penalties, and context compression. Across these runs, we identify hub over- load as a sharp residual failure mode: when high- centrality distractor nodes dominate the retrieved context, performance collapses. A simple hub- penalty mitigation reduces hub overload but also hurts aggregate retrieval quality, showing that the failure cannot be repaired by naive post-hoc score adjustment alone