Abstract
Adaptive retrieval promises to enhance the robustness of knowledge-graph question answering by allowing a controller to search, inspect neighborhoods, revise actions, and stop when evidence is sufficient. We explore this premise by introducing TraceBound, a lightweight profile- and trace-conditioned diagnostic protocol for an ARK-style retriever on text-rich knowledge graphs.
TraceBound exposes a compact query profile before retrieval, issues short trace hints after observable failure symptoms, and logs trajectory counters while keeping graph data, tools, gold labels, and ranking metrics fixed. Across STaRK validation and held-out subsets, the added conditioning improves inspectability but consistently reduces retrieval quality under open-weight controllers. Paired trajectory analysis localizes the degradation to repeated calls, zero-result calls, and misallocated exploration budget, while stricter interaction budgets shorten trajectories without repairing the policy. The result diagnoses the common failure mode that "thinking before retrieval" must be evaluated as a control problem over action selection, not merely as a prompt-format change.
Blogger's Review: The study of TraceBound reveals that the effectiveness of control strategies in knowledge graph retrieval is more crucial than merely adjusting prompt formats. Through precise trajectory analysis, we can identify the specific reasons for the decline in retrieval quality, providing an important theoretical foundation for future retrieval optimizations.