Holographic Memory for Zero-Shot Compositional Reasoning in Knowledge Graphs: A Mechanistic Study of Where and Why It Fails

Short summary

Researchers tested whether Holographic Reduced Representations could enable knowledge graphs to answer multi-hop questions never seen during training. While HRRs matched baseline performance on single-hop queries, they failed at compositional reasoning—accuracy remained at chance level. The bottleneck isn't the bind-unbind mechanism, but retrieval capacity under superposition.

• •Holographic Reduced Representations achieve strong single-hop retrieval but completely fail at zero-shot compositional reasoning across all tested conditions
• •The failure mechanism is fundamental capacity limitations under superposition, not the theoretical bind-unbind algebra or cleanup design
• •Fixing compositional reasoning requires improving memory capacity, not redesigning the mechanism

Generated with AI, which can make mistakes.