Structured knowledge retrieval drops hand-written query-rule pairs: a routing agent and its retriever train in alternation

Structured knowledge retrieval can now decide for each query which reasoning rule to apply, then search the knowledge base for evidence under that rule, without hand-written query-rule pairs. Earlier methods either retrieved without constraints or relied on predefined query-rule supervision, such as author-paper rules for scientific discovery and anatomy-drug rules for precision medicine, and such supervision is rarely available in practice. Experiments on STaRK-QA datasets spanning scientific discovery, biomedical reasoning, and e-commerce knowledge management, with ablations showing a benefit from co-adapting the router and retriever.

A routing agent first selects a rule, and the retriever corresponding to that rule traverses the knowledge base and ranks evidence satisfying the rule's constraints. Because routing and retrieval are interdependent, CoRE alternates between reinforcement learning for the routing policy and contrastive learning for the rule-constrained retriever. The selected rules supervise retriever training while retrieval performance rewards the routing policy, a closed loop that lets both components co-adapt without explicit query-rule supervision.

This report covers what the summary states: CoRE targets structured knowledge retrieval and is evaluated on STaRK-QA datasets across scientific discovery, biomedical reasoning, and e-commerce knowledge management. It suits researchers and engineers working on knowledge retrieval, retrieval-augmented generation, or rule-constrained reasoning. The summary says generation-based rule routing transfers across datasets and can move from multi-domain pretraining to target-specific fine-tuning, without giving the transfer setup.

The summary states only that experiments on STaRK-QA datasets demonstrate effectiveness, without specific metrics or improvement margins. The ablations show a benefit from co-adaptation but not the size of each component's separate contribution. Rule-usage analysis is described as informing automatic rule refinement and discovery, which is a direction rather than a verified conclusion.