acceptodds
Under review as a conference paper at ICLR 2027

CoverKG: Coverage and Compactness Dual Rewards for Knowledge Graph Extraction

Abstract

Knowledge-graph-based retrieval-augmented generation relies on an LLM extractor to build a graph index, the quality of which directly governs retrieval effectiveness and storage efficiency. Yet existing methods supervise the extractor only through downstream QA accuracy. This sparse, end-to-end signal cannot diagnose whether failures arise from insufficient semantic coverage or from redundant extractions that introduce noise into retrieval. We introduce CoverKG, a framework that characterizes index quality through two complementary objectives, namely coverage and compactness, and converts this characterization into a construction-time training signal that can be computed for every source chunk and used to optimize the extractor with reinforcement learning. Coverage aligns the source sentences with the extracted graph elements to pinpoint missing semantics at the sentence level. Compactness evaluates each element's marginal contribution relative to the other elements in the same extraction, directly penalizing redundancy. A composite reward balancing both objectives drives policy optimization. The framework is readily applicable to the graph construction stage of various graph-based RAG systems. Experiments across multiple graph-based RAG pipelines and QA benchmarks show that, even under the same downstream evaluation protocol as prior work, extractors trained with CoverKG consistently achieve higher accuracy with more compact graph indices, demonstrating that per-chunk, dual-objective supervision at index construction time yields both more effective retrieval and more efficient storage. Our code is available at https://anonymous.4open.science/r/anonymous_repo-1357.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Graph-based Retrieval-Augmented Generation, Knowledge Graph Construction, LLM Post-training, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6X7xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CoverKG, which trains a knowledge graph extractor using rewards for coverage and compactness. Coverage measures the similarity between source sentences and extracted descriptions, while compactness is computed from the coverage lost when each element is removed. The two rewards are combined to train Qwen3-8B with GRPO without QA labels. Experiments with GraphRAG, LightRAG, and HippoRAG2 show improvements in 12 of 15 QA configurations and substantial reductions in index size.

Critical strengths
  1. The reward can be computed directly from source text and candidate extractions, without QA annotations or running a downstream QA pipeline during each training rollout.

  2. The experiments cover three different retrieval backends and several benchmarks. The method substantially reduces index size while improving QA performance in many settings, suggesting practical value for graph construction.

Critical weaknesses
  1. The compactness reward can increase after adding a redundant element, contradicting lines 244–248. It exposes an incentive inconsistent with the claimed redundancy penalty, although it does not establish that the trained model exploited this behavior. Consider three equally weighted sentences and five extracted elements. Their similarities are:

    Sentence 1: (0.80, 0, 0, 0, 0)
    Sentence 2: (0, 0.80, 0, 0.79, 0)
    Sentence 3: (0, 0, 0.80, 0, 0.79)

    With tau = 0.63, coverage is 0.80. Removing each element in turn gives marginal contributions of approximately (0.27, 0.0033, 0.0033, 0, 0). Now add a sixth element with the same description embedding as the first. Coverage remains unchanged, but the first and sixth elements can replace each other, so removing either causes no loss of coverage. The marginal contributions become (0, 0.0033, 0.0033, 0, 0, 0).

    Although this adds no coverage and makes the extraction more redundant, the remaining positive contributions are now equal. Their participation ratio therefore increases enough to outweigh the penalty from the larger element count: compactness rises from 0.21 to 0.33, and the reward at beta = 1 rises from 0.33 to 0.47.

  2. The experiments do not sufficiently support the benefit of the proposed reward. The paper compares Base, SFT, and CoverKG, but omits directly relevant methods such as AutoGraph-R1. Moreover, the ablation studies are limited to GraphRAG, while LightRAG and HippoRAG2 backends are not evaluated.

Questions for authors
  1. How do you reconcile the counterexample with lines 244–248? Does the same behavior occur with embeddings from the actual model, and how does it affect training?

  2. How does CoverKG compare with AutoGraph-R1 and other relevant methods? Do the ablation results also hold for LightRAG and HippoRAG2?

Reviewer_V7f8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoverKG, a reinforcement learning framework for knowledge graph extraction in graph-based RAG systems. Its construction-time reward requires neither downstream QA labels nor gold graphs. Coverage aggregates thresholded maximum semantic similarities between source sentences and extracted element descriptions, while compactness applies a normalized participation ratio to leave-one-out coverage marginals. Following supervised initialization, the extractor is optimized using GRPO. The main evaluation spans GraphRAG, LightRAG, and HippoRAG2 on five QA benchmarks, with additional medical-domain experiments on GraphRAG-Bench. CoverKG outperforms Base in 12 of 15 backend–benchmark combinations and substantially reduces graph size and serialized storage footprint.

Critical strengths
  1. The paper addresses a practically important supervision problem. Per-chunk rewards during graph construction reduce dependence on downstream QA labels and gold graphs. This approach is promising for adapting knowledge graph extraction to domains with limited supervision.
  2. The coverage formulation and marginal computation are clear. Thresholded semantic alignment provides an interpretable sentence-level coverage proxy. The top-two alignment procedure efficiently computes all leave-one-out marginals without separately recomputing coverage for every removed element.
  3. Cross-backend evaluation provides useful empirical evidence. The five-benchmark evaluation spans three distinct graph-based RAG backends, with additional medical-domain experiments. Improvements in 12 of 15 configurations, together with substantial graph compression and storage reductions, support the approach’s potential practical value.
Critical weaknesses
  1. The compactness reward does not always penalize redundancy as claimed. Section 4.2 claims that adding elements without increasing coverage cannot raise the reward. With sentence weights (0.9, 0.1) and an initial identity similarity matrix, adding an element with similarities (8/9, 0) leaves coverage unchanged but changes the marginals to (0.1, 0.1, 0). Compactness rises from 0.610 to 0.667, and the reward from approximately 0.758 to 0.800 for β = 1, ignoring the small stabilizer. This contradicts the stated claim in Section 4.2 and warrants correcting the formulation or its interpretation and testing near-duplicate elements.
  2. Semantic coverage and compression do not establish information preservation. Maximum cosine alignment does not directly verify relation correctness, direction, or multiple facts within a sentence. In Table 2, GraphRAG’s medical-domain edge count falls from 3,168 to 81. Downstream gains and the Appendix E case do not establish that only redundant information was removed. Fact-level precision and recall, relation accuracy, and multi-hop evidence-path preservation would provide stronger support.
  3. The specific benefits of the reward design remain insufficiently isolated. Table 4’s component ablations support the combined reward but do not establish the advantage of participation ratio over simpler length or element-count penalties, or size-matched deduplication and pruning. The ablations do not report corresponding graph-size or information-retention measurements. Further training, indexing, and retrieval cost measurements would clarify overall efficiency beyond the demonstrated storage savings.
  4. Robustness and generalization evidence remain limited. Some QA gains are small, but confidence intervals and repeated GRPO training runs are missing. Although QA labels are unused, document-level overlap between training and evaluation remains unclear for HotpotQA and MuSiQue. The reward also relies on Qwen3-Embedding-4B and an Otsu-selected threshold of τ = 0.63, while sensitivity analysis mainly varies β. Clearer data partitions and calibration details, alongside encoder and threshold sensitivity tests, would strengthen the evidence.
Questions for authors

See weaknesses.

Reviewer_frzPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the task of graph construction in graphrag. Specifically, the paper proposes the coverkg method. The purpose of this method is to train an LLM for constructing the graph. In this method, it first trains an LLM through regular SFT to let it learn the basic output format. Then, based on SFT trained LLM, additional GRPO training is performed on the model based on two automatic reward functions. One is to focus on evaluating the coverage of the extracted graph to determine whether every sentence that exists in the original text is present in the graph through semantic similarity. The other is to calculate the compactness score by deleting items that exist in the graph to see if it will lead to a decrease in coverage. The authors plug the model trained using their method directly into the existing graphrag frameworks. Experimental results show that the method proposed by the authors can significantly reduce the size of the graph and improve performance in some settings.

Critical strengths

I think a critical strength of this paper is the design of its reward mechanism. The reward functions in this paper use simple semantic similarity to compare the extracted entities and relations with the original text and do not require additional human annotation.

Critical weaknesses

I think the method presented in this paper has two critical weaknesses. Firstly, while I mentioned above that its reward function calculation does not require additional manual annotation, the scoring rule of the proposed method does not consider the graph structure itself. If the textual description of a node contains some meaning from a sentence in the text, then the relation between that node may be omitted by the LLM. Because the proposed method is designed so that the meaning from the text only needs to appear once in the graph. This could lead to a graph missing many necessary relations, or even degenerating into a simple collection of disconnected nodes. Secondly, I also think this paper lacks comparisons with other LLM models trained beyond simple SFT to demonstrate the effectiveness of its proposed method.

Questions for authors

Besides comparing the proposed method with the base and sft models, have you compared it with any other methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.