TL;DR for operators

An enterprise retrieval agent can rank the available records accurately yet fail because it searched through the wrong relationships. Co-Adaptive Rule Routing and Constraining Retrieval (CoRE), introduced by Qi and colleagues,1 trains the choice of retrieval path alongside the components that rank evidence. Across three structured-retrieval benchmarks, its strongest configuration achieves 54.47 average Hits@1. Removing learned routing adaptation reduces that score to 16.00, while replacing learned retrievers with fixed scoring reduces it to 48.08. These ablations identify routing as a major bottleneck in the evaluated architecture, rather than proving its universal superiority over ranking. For enterprise knowledge systems, the research suggests measuring path-selection failures separately before investing in stronger retrievers. Deployment reliability, access-control enforcement, and downstream business value remain untested.

An accurate ranker can still search the wrong records

Consider a research assistant asked to identify publications associated with a particular institution. Its knowledge graph connects institutions, authors, papers, and research topics.

The assistant could follow the institution-to-author-to-paper relationship. Alternatively, it might choose another plausible relationship that leads to a different set of records. Both paths may be permitted by the database schema, but only one may reach the evidence needed to answer the question.

Once the wrong path has been selected, improving the ranking model cannot recover evidence excluded from the candidate set.

This creates two different engineering problems: selecting a productive route through structured knowledge and ranking the evidence reachable through that route.

CoRE investigates whether these decisions should be learned together. Its central observation is that the first decision can substantially constrain the effectiveness of the second.

CoRE trains the route and the retriever together

CoRE begins with a predefined inventory of schema-valid reasoning rules. A rule specifies an allowed sequence of relationships, such as:

Institution → Author → Paper

The architecture has two principal components.

The routing agent selects a rule based on the natural-language query. The rule-constrained retriever traverses the knowledge graph according to that rule and ranks the reachable evidence.

The retriever uses textual similarity signals from entities along each traversal path, combining BM25 features with a learned representation of the selected rule. Its implementation retains up to 200 candidates per query–rule pair.

The innovation is how the two components learn from each other.

During training, the router assigns queries to rules, giving specialized retrievers examples on which to improve their rankings. The system then evaluates how effectively each rule retrieves known evidence. Better-ranked ground-truth evidence produces a larger reciprocal-rank reward.

Those rewards inform subsequent routing updates. Updated routing assignments, in turn, change the examples used to train the retrievers. The procedure alternates for three rounds.

Importantly, CoRE does not require labeled query–rule assignments, but it still requires query–evidence supervision. Nor does it invent unrestricted graph relationships: its rule choices come from the available schema-valid inventory.

At inference, it selects one rule and retrieves evidence through that rule, rather than combining results from every possible path.

Removing router adaptation causes the larger performance loss

The clearest evidence comes from the component ablations in Section 5.3, Table 2.

Configuration Average Hits@1 Average MRR
Full CoRE-BERT-C 54.47 65.97
Without router adaptation: unadapted BERT 16.00 21.01
Without router adaptation: GPT-5-mini 19.69 23.75
Without retriever adaptation: fixed BM25 traversal 48.08 58.87

Source: Section 5.3, Table 2. Values are reported on a 0–100 scale across Amazon, MAG, and PRIME.

With the unadapted BERT router, average Hits@1 falls by 38.47 percentage points despite continued retriever training. With fixed BM25-traversal scoring instead of learned retrievers, the decline is 6.39 points, while routing continues to adapt.

The asymmetry matters because it identifies a specific failure dependency. A specialized retriever cannot compensate reliably when its router assigns queries to paths that fail to expose useful evidence.

The retriever ablation nevertheless shows that ranking contributes independently. Learned path-level scoring improves results beyond structural filtering, particularly on the biomedical PRIME dataset.

These are component-removal tests under particular training configurations. They do not establish that routing always contributes more than ranking, or that the measured performance differences can be attributed to either component independently of how the remaining system was trained.

The benchmark advantage does not extend to every dataset

The main comparison in Section 5.2, Table 1 evaluates CoRE against textual, graph-based, hybrid, and agentic retrieval methods on three STaRK-QA datasets.

Together, the datasets contain 33,890 queries across e-commerce, academic publications, and biomedical knowledge graphs. Their candidate inventories range from six rules on Amazon to 65 on PRIME.

CoRE-BERT-C achieves average Hits@1 of 54.47, compared with 48.93 for the hybrid MoR baseline. Average mean reciprocal rank rises from 58.77 to 65.97. These are improvements of 5.54 and 7.20 percentage points, respectively.

But the advantage is not uniform.

On Amazon, KAR achieves 54.20 Hits@1 against CoRE-BERT-C’s 50.71. On MAG, HYBGRAG reports 65.40 against CoRE’s 63.37. Some methods also lack Amazon results, making their published cross-dataset averages less directly comparable.

Within CoRE, the BERT classification router also outperforms the more computationally demanding generative alternative on average. CoRE-LLM-G achieves 48.31 Hits@1 versus 54.47 for CoRE-BERT-C.

The result argues against automatically assigning a larger generative language model to every retrieval-routing decision.

Plausible reasoning paths are not necessarily effective retrieval paths

Why not simply ask a capable LLM to choose the appropriate graph relationship?

Appendix C provides a diagnostic answer. Its routing-failure analysis finds cases where GPT-5 selects semantically plausible rules that nevertheless retrieve evidence poorly. The analysis identifies both misalignment between semantic reasoning and retriever effectiveness, and incorrect grounding in graph relationships.

A path can sound appropriate without reaching the right records or producing a ranking that places relevant evidence near the top.

CoRE therefore uses actual retrieval outcomes as training feedback rather than treating semantic plausibility as sufficient supervision.

Other experiments examine the broader behavior of this approach. The MAG optimization trace shows Hits@1 rising from approximately 21% to 58% after router training, with a subsequent retriever update pushing it beyond 62%. This illustrates the alternating process; it is not a formal convergence guarantee.

Cross-dataset tests also find that trained generative routers can improve over an unadapted Qwen router on other domains. Shared multi-domain training followed by target-specific fine-tuning strengthens the generative routing results. However, these comparisons retain trained target-domain retrievers and dataset-specific schemas. They demonstrate partial transfer of routing behavior, not transfer of an entire retrieval system without adaptation.

Enterprise retrieval needs separate failure measurements

For an enterprise knowledge-platform team deciding whether to upgrade embeddings, retrievers, or LLM-based orchestration, Cognaptus draws a specific architectural inference.

A retrieval evaluation should distinguish three questions:

  1. Was the route appropriate? Did the agent select relationships capable of reaching the evidence required by the query?
  2. Was the evidence reachable? Did the selected path, graph coverage, and candidate-generation procedure actually include the relevant records?
  3. Was the evidence ranked correctly? Once retrieved, did the scoring component place relevant records sufficiently high for the downstream task?

These measurements support different engineering decisions.

If relevant records are consistently unreachable, retriever tuning alone is unlikely to resolve the problem. If the records are reachable but poorly ranked, improving relevance scoring becomes more defensible. If some necessary relationships do not exist in the graph, neither routing nor ranking adaptation can create those missing facts.

A customer-support knowledge system, for example, could evaluate routing between products, issues, and troubleshooting records against verified query–evidence examples. Its operator could then allocate effort according to observed failure categories, while separately enforcing data-access permissions.

This is a proposed evaluation workflow informed by CoRE, not an enterprise deployment demonstrated by the paper.

Better retrieval rankings are not yet better business decisions

Several boundaries constrain what can be concluded.

CoRE depends on predefined reasoning-rule inventories. Its single-rule inference can exclude relevant evidence from alternative routes, while the 200-candidate limit introduces another possible source of omission. The benchmark does not establish robustness to incomplete schemas, changing relationships, corrupted graphs, or adversarial inputs.

The principal comparisons also do not report confidence intervals or repeated-seed variability. Reproduction requires attention to a methodological distinction: the main description presents reward-relative classifier optimization, while Appendix A describes hard best-rule targets for BERT-C; Appendix D separately compares training objectives.

Efficiency evidence is similarly narrower than a production cost analysis. Appendix F compares inference time on 1,000 MAG questions and indicates lower routing overhead for classifier-based systems, but does not establish total enterprise operating costs.

Most fundamentally, the evaluated outcomes concern evidence retrieval rankings. They do not measure whether an assistant produces correct final answers, respects permissions, makes sound clinical judgments, or improves financial performance.

The next retrieval upgrade may be a routing evaluation

CoRE’s strongest contribution is showing that structured retrieval quality depends on the learned interaction between path selection and evidence ranking.

Its ablations justify treating the router as a separately measurable component, while the benchmark comparison provides evidence that joint adaptation can improve retrieval results within established graph schemas.

For organizations building agents over connected enterprise records, the next improvement may begin with identifying where relevant evidence becomes unreachable—before spending more compute on ranking the records that remain.

Cognaptus: Automate the Present, Incubate the Future.


  1. Zhisheng Qi and Haoyu Han and Jiaxi Li and Bing He and Huiyuan Chen and Franck Dernoncourt and Ryan A. Rossi and Mahantesh M. Halappanavar and Jin Lu and Yu Zhang and Tianming Liu and Yu Wang (2026). Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval. arXiv:2610.14465. https://arxiv.org/abs/2610.14465 ↩︎