A new arXiv preprint, arXiv:2610.14465, addresses the challenge of retrieving task-relevant evidence from structured knowledge sources. The authors note that existing methods typically fall into two camps: unconstrained retrieval, which may pull in irrelevant information, and approaches that depend on predefined query-rule supervision, such as fixed author-paper rules. Both have limitations when dealing with complex or varied queries.

The paper proposes a co-adaptive mechanism that simultaneously routes and constrains rules during retrieval. Rather than relying on static rules or fully open-ended search, the method adapts its rule application based on the query and the structure of the knowledge source. This allows the retrieval process to stay focused on task-relevant evidence while remaining flexible enough to handle different query types.

Because the abstract is truncated, full implementation details and experimental results are not yet available from the source. The contribution, as described, is a conceptual shift toward co-adaptation in rule-based retrieval for structured knowledge, which could improve downstream tasks that depend on accurate evidence selection.