STEM: Structure-Tracing Evidence Mining for Knowledge Graphs-Driven Retrieval-Augmented Generation

The core bet STEM makes is that treating retrieval as a graph search problem, constrained by schema structure, is more reliable than embedding-similarity lookups that ignore relational topology. That framing matters because most RAG criticism focuses on chunking and context length, not on the structural mismatch between how queries are posed and how knowledge is actually organized in graphs.

This connects most directly to IG-Search, covered here in mid-April, which attacked a related failure mode from a different angle: rather than fixing how retrieval maps to structure, IG-Search used reinforcement learning to reward queries that actually improve model confidence. The two approaches are complementary in theory but competing in practice, since both claim to address multi-hop reasoning degradation without agreeing on where the root cause lives. DiscoTrace, also from mid-April coverage, adds a third lens by showing that LLMs systematically favor breadth over selectivity when constructing answers, which is exactly the behavior that poor graph retrieval tends to amplify.

The real test is whether STEM's globally-aware node anchoring holds on datasets with sparse or noisy schema coverage, not just well-formed benchmarks like WebQSP. If ablations on schema-incomplete graphs show significant accuracy drops, the approach is more brittle than the framing suggests.