Train the Store, Audit the Facts: What KBevo Changes About Retrieval

TL;DR for operators

Most retrieval systems treat the external knowledge layer as fixed infrastructure: build an index or graph, then improve the model, retriever, or reranker around it. KBevo tests a different decision. It lets downstream answer quality influence how the knowledge store itself is constructed.

That changes what can be optimized. After reinforcement learning, average exact match rises from 36.2 to 41.3 for the 1.7B model and from 36.8 to 46.6 for the 4B model across four QA benchmarks. The resulting structured store can also be edited and reused without retraining the language model.

The constraint is equally important. The learned knowledge bases become more useful for answering questions, but their stored triplets become less faithful by the paper’s diagnostics. An organization considering this design therefore needs two evaluations: whether the knowledge layer improves downstream task performance, and whether its contents are sufficiently verified for the consequences attached to those facts.

The knowledge store does not have to remain fixed

In a typical retrieval pipeline, changing facts create an awkward division of labor. The external store is updated through a separate indexing or knowledge-engineering process, while model optimization concentrates on retrieving or reasoning over whatever the store contains.

That separation is manageable when the store is only a source of passages. It becomes more consequential when the system repeatedly answers related questions and the external layer is expected to function as durable operational memory. A store can contain plenty of information yet organize it poorly for the questions the system actually needs to answer.

Co-Evolving Structured Knowledge and Reasoning in Language Models introduces KBevo to test whether the construction of that external knowledge layer can itself be optimized from downstream outcomes.¹ Instead of retrieving unstructured passages, the system builds an editable collection of entity-relation-value facts and learns to answer questions through targeted lookups into that structure.

The key change is not merely that the store is structured. It is that answering performance feeds back into how the structure is learned.

Downstream answers become credit for the knowledge base

KBevo uses one shared Qwen3 policy for two phases. First, the model extracts structured triplets from source material. It then answers questions by retrieving from those triplets.

During reinforcement learning, the system samples four candidate knowledge bases for a question and eight answer rollouts for each candidate. The answer F1 scores are averaged, and that average becomes the reward assigned to the candidate knowledge base that enabled those answers.

In plain language, a knowledge base receives credit when it makes correct answers easier to produce.

That mechanism is what makes the paper’s use of “co-evolution” concrete. The reasoning policy and the external store are not optimized as independent components. The same downstream result alters both what the model learns to store and how it learns to use the stored structure.

The main QA results show that this feedback matters. Across HotpotQA, MuSiQue, 2WikiMultiHopQA, and PopQA, reinforcement learning improves KBevo over its supervised-fine-tuned starting point at both tested scales.

These are the paper’s main benchmark results, not an appendix sensitivity test. They support the narrower claim that downstream-coupled training improves this particular structured retrieval-and-reasoning system.

They do not show that structured retrieval is categorically more accurate than unstructured search.

Accuracy alone does not justify the architecture

The matched Search-R1 comparison is useful precisely because KBevo does not dominate it. Search-R1 reaches 42.4 average exact match at 1.7B versus KBevo-GRPO’s 41.3, and 49.0 versus 46.6 at 4B.

The per-benchmark pattern is mixed. KBevo-GRPO performs better on MuSiQue and 2Wiki at both scales, while Search-R1 remains ahead on HotpotQA and PopQA.

That moves the deployment decision away from a simple leaderboard question. If the only requirement is aggregate QA accuracy under these benchmark conditions, the paper does not establish KBevo as the superior choice. Its additional value comes from properties of the external store.

One ablation makes that distinction clearer. With the 1.7B reasoning policy held fixed, the learned KBevo knowledge base reaches 35.6 average EM on the tested subsets, compared with 26.2 for EDC and 24.6 for AutoSchemaKG. A Gemini-generated knowledge base reaches 37.2. The likely purpose of this ablation is to isolate knowledge-base quality from changes in the reasoning model. It suggests that downstream-coupled construction produces a store substantially more useful to the fixed KBevo reasoner than the two static KG baselines, while also showing that KBevo has not exhausted the attainable quality of the store itself.

For operators, that is a more informative result than claiming a universal retrieval win.

Editability is the property that changes the operating model

A structured external store creates a separate update surface from the model parameters. The paper tests that property with counterfactual edits from ConFiQA-MR: facts in the external knowledge can be changed, and KBevo can then use those edited facts in multi-hop reasoning without retraining the model.

This is direct paper evidence for controllability, not evidence about production update latency, governance cost, or enterprise-scale maintenance.

Cognaptus inference: the design is most relevant when the same knowledge corpus supports repeated queries and some facts change more often than the model should be retrained. In that setting, offline knowledge construction can be amortized across requests, while selected facts remain independently editable.

The affected decision is whether to keep the external knowledge layer as passive retrieval infrastructure or make it a task-optimized component. The case is strongest where updateability and reuse matter alongside answer quality. The paper does not establish the same benefit for open-ended assistants, rapidly changing streams, or tasks outside knowledge-intensive QA.

Better task utility produces a worse faithfulness profile

The most important qualification appears in the knowledge-quality diagnostics.

After reinforcement learning, KBevo’s knowledge bases improve in coverage and connectivity. But triplet-level correctness moves in the opposite direction. In the paper’s LLM-judged 100-sample analysis, valid triplets fall from 92.3% to 85.0% for the 1.7B model and from 96.0% to 84.3% for the 4B model. Hallucinated and soundness-problem rates both rise.

This appendix analysis is a faithfulness diagnostic, not a second benchmark claim. Some Section 5.3 and Appendix C diagnostics use an earlier top-$k=1$ checkpoint rather than the final top-$k=4$ configuration, and the correctness labels rely on an LLM judge. The percentages therefore should not be treated as directly interchangeable with the final QA results.

The direction of the trade-off still matters. The reward is based on downstream answers, and only retrieved triplets directly influence that reward. The optimization process therefore has no reason to guarantee that every stored fact becomes globally faithful.

The paper further reports that filtering unfaithful triplets slightly improves exact match. That weakens a tempting alternative explanation: the QA gains are not simply being purchased by storing useful hallucinations.

For any system treating external knowledge as operational memory, task reward and knowledge integrity are separate control problems.

The method also has a capability prerequisite

KBevo is not a demonstration that reinforcement learning can discover structured database reasoning from a base model without preparation.

Direct RL from the base model collapses in the reported experiments. Supervised fine-tuning is needed to establish the basic two-phase behavior first, although an early SFT checkpoint is sufficient to bootstrap later reinforcement learning.

This ablation defines a deployment dependency: organizations cannot assume that outcome rewards alone will create the required interaction protocol between extraction, lookup, and reasoning. The policy first needs enough competence to use the structured interface at all.

Treat the learned store as optimized memory, not trusted memory

KBevo provides strong benchmark evidence for a specific architectural shift: external knowledge construction can be coupled to downstream task outcomes rather than treated as fixed preprocessing.

For systems serving repeated knowledge-intensive requests, that opens a credible design option. Optimize the store for the work the model must perform, preserve it outside the model so facts remain editable, and reuse it across requests.

But the paper also supplies the reason not to collapse utility and trust into one metric. A knowledge base can become more reachable, more connected, and more useful for answering questions while becoming less faithful at the triplet level.

The deployment rule follows directly from that evidence: optimize the external knowledge layer for task performance if the workload benefits from it, but verify the learned contents through a separate provenance or validation process before allowing them to function as trusted operational memory.

Cognaptus: Automate the Present, Incubate the Future.