Are Hypervectors Enough? Single-Call LLM Reasoning over Knowledge Graphs

When large language models (LLMs) are grounded in knowledge graphs (KGs), they need to reason over symbolic relation paths to answer queries. Many current systems embed and score these paths with neural encoders, which raises compute cost, increases latency, and makes decisions harder to interpret. We present PathHD, a light, low-cost, and transparent KG reasoning framework that replaces neural path scoring with hyperdimensional computing (HDC). PathHD encodes relation paths as block-diagonal GHRR hypervectors, retrieves candidates with fast cosine similarity and Top-K pruning, and uses a single LLM call to produce the final answer while citing the supporting paths. Technically, PathHD provides a binding operator that preserves the step-by-step relation order in the path and is invertible for inspection, a calibrated similarity that yields robust retrieval, and a single-shot decision step that removes per-path LLM scoring while keeping interpretability. On WebQSP, CWQ, and GrailQA, PathHD (1) matches or exceeds strong neural baselines on Hits@1 with one LLM call per query, (2) reduces end-to-end latency by $40$–$60$% and GPU memory by $3$–$5$$\times$ through encoder-free retrieval, and (3) returns faithful, path-grounded rationales that help error diagnosis and controllability. These results show that HDC is a practical basis for efficient KG-grounded LLM reasoning, balancing accuracy, efficiency, and interpretability.