acceptodds
Under review as a conference paper at ICLR 2027

Zero-LLM Trajectory Search over Line-Lifted Knowledge Graphs: Internalizing Query Decomposition for Multi-Hop RAG

Abstract

Retrieval-augmented generation (RAG) has become a cornerstone of LLM applications, underpinning vertical-domain knowledge construction and agent memory. As applications demand complex relational reasoning, retrieval has turned to graphs, yet graph-based RAG still relies almost entirely on LLM calls, one or more per query, making latency and cost a deployment bottleneck, as if precise retrieval required LLM guidance. We propose LiftRAG, which represents every extracted relation as a graph vertex carrying its evidence-sentence embedding, turning the knowledge graph into its line graph. Retrieval is then a multi-round beam search over entity–relation trajectories, where each step decides the next hop from the frontier reached so far and query decomposition emerges from the search itself, LLM-free. On the entity-grounded test sets of three multi-hop benchmarks, LiftRAG leads the published baselines on all four metrics. The accuracy margins are 5.7–10.0 points over LightRAG, which shares our extraction and index, and 0.9–3.9 over the strongest baseline, HippoRAG 2, at 13–25× lower retrieval latency in API deployment (per-call latency measured on our own LLM relay). Moreover, 7–20% of LLM-decomposed sub-question trajectories fall outside the coverage of a single retrieval by LiftRAG. This work establishes Zero-LLM Decomposition, query decomposition performed by the retrieval process itself without any LLM call. The part of LLM decomposition it does not cover defines the open problem of fusing the two.

Primary area:
foundation or frontier models, including LLMs
Keywords:
retrieval-augmented generation, knowledge graphs, multi-hop question answering, query decomposition, LLM-free retrieval

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vyC5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LiftRAG, which models every extracted relation as a graph vertex carrying its evidence-sentence embedding, turning the knowledge graph into its line graph. In addition, this paper proposes a multi-round beam search over entity-relation trajectories to replace per-query LLM guidance with fixed constants. This paper also conducts extensive experiments to show the effectiveness of the proposed method.

Critical strengths
  1. The idea behind transforming KGs into line graph is interesting.
  2. This paper proposes a training-free method, which may reduce the cost for solving questions.
  3. Extensive experiments demonstrate the effectiveness of the proposed method.
Critical weaknesses
  1. The technique novelty behind this paper appears limited, as line graph transformation and multi-round beam search may borrowed from existing methods. It would be beneficial to clarify the difference.
  2. It may lack some recent state-of-the-art baseline methods for multi-hop question answering, such as LinearRAG [1], GFM-RAG [2].
  3. It would be beneficial to include a comprehensive comparison of graph construction costs, inference time to show the advantage of training-free method.
  4. It would be beneficial to polish the manuscripts to improve the readability. For example, put the critical experimental results in the main manuscripts instead of placing the conclusion and increase the font size in the figure 2.

[1] Zhuang, Luyao, et al. "Linearrag: Linear graph retrieval augmented generation on large-scale corpora." International Conference on Learning Representations. Vol. 2026. 2026. [2] Luo, Linhao, et al. "GFM-RAG: graph foundation model for retrieval augmented generation." Advances in Neural Information Processing Systems 38 (2026): 36371-36405.

Questions for authors

Please refer to the weaknesses of the paper. Main concerns involve the limited technique novelty and presentation of manuscripts.

3 comments

We would like to share a set of additional experiments that we ran in response to a request from another reviewer, because we believe they are relevant to your assessment as well.

The reviewer asked whether our gains come from the graph mechanism or simply from searching a larger candidate pool. We therefore matched the candidate-pool size in characters across methods and re-ran the comparison, and we also added dense top-100 retrieval followed by a strong cross-encoder reranker under the same 8,000-character budget. At matched pool size our pool-level recall is at parity with dense retrieval and within two points of HippoRAG 2, while our context recall and our conversion from pool to context are the highest on all three datasets. Against dense retrieval with the reranker we lead by 8.2 and 5.1 accuracy points on 2WikiMQA and MuSiQue and tie on HotpotQA, where the dense pool is 1.2 to 2.5 times ours. A coverage analysis further shows that the two kinds of pool reach different evidence, which is the complementarity the paper identifies as future work.

We believe these results further support the novelty and potential of the method, and they may affect your view of it. The full tables (R1 to R4) are in our response to Reviewer Ap7P. We hope this resolves the remaining concerns at once, and we are happy to clarify any number.

Thank you for your response. It mainly addresses my concerns, and I will increase the rating to 3.

We are very glad that our work has been recognized by you. Wish you all the best in your academic work.

Reviewer_Ap7POpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LiftRAG, a training-free GraphRAG pipeline that leverages an entity-relation graph to enhance retrieval. Specifically, LiftRAG first retrieves relevant passages, entities, and relations using the original query, and then iteratively expands the retrieved entities and relations over TT rounds. After constructing a candidate pool of triples and passages, LiftRAG aggregates the retrieved evidence and feeds it into the reader for answer generation. Experiments on three multi-hop QA benchmarks demonstrate consistent improvements over the baselines.

Critical strengths
  1. The paper is well written, clearly structured, and easy to follow.

  2. The motivation is reasonable and well justified.

  3. The proposed method is lightweight and achieves promising empirical results without requiring additional training.

Critical weaknesses
  1. Missing references to relevant prior work. The paper overlooks several closely related studies. For example, LinearRAG [1] and E²GraphRAG [2] are also graph-based RAG methods that do not require LLM calls during retrieval. These methods should be discussed and compared against.

  2. Insufficient details about baseline configurations and potential fairness concerns. The paper does not disclose the hyperparameter settings of the baselines, making it difficult to assess the fairness of the comparisons. In particular, I am concerned that part of the observed improvements may stem from LiftRAG's larger candidate retrieval pool rather than its proposed graph-based expansion mechanism. Since LiftRAG performs multiple rounds of retrieval and expansion before selecting evidence under a fixed context budget, a fairer comparison would allow conventional retrieval baselines to retrieve a similarly large candidate pool followed by reranking and evidence selection.

  3. Limited methodological novelty. The proposed method appears to combine several well-established techniques, including multi-channel retrieval, incidence graph construction, query-guided graph traversal, beam search, candidate accumulation with score normalization, and evidence fusion with budget allocation. Although their integration may be practically useful, the paper does not sufficiently clarify which components constitute substantive methodological innovations beyond existing approaches.

[1] LinearRAG: Linear Graph Retrieval Augmented Generation on Large-scale Corpora, Zhuang et. al, ICLR 2026.

[2] E$^2$GraphRAG: Streamlining Graph-based RAG for High Efficiency and Effectiveness, Zhao et.al, Arxiv preprint.

Questions for authors
  1. Could the authors provide a complete list of hyperparameter settings for all baselines, particularly retrieval depth, candidate pool size, reranking configuration, and final context budget? Additionally, could the authors compare LiftRAG against baselines equipped with larger retrieval pools and reranking under the same final context budget?

  2. Could the authors include a direct comparison with LinearRAG [1] to better establish the effectiveness of LiftRAG relative to closely related training-free graph-based retrieval approaches?

  3. Could the authors include a comparison with retrieval-then-rerank vector RAG with a higher recall kk?

Reviewer_Rb3VOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LiftRAG, a graph-based retrieval framework that aims to eliminate retrieval-time LLM calls for multi-hop question answering. The method transforms relation instances into retrievable vertices with evidence-sentence embeddings and performs multi-round trajectory search using beam search. This allows the retrieval process to explore multi-hop relations without explicit LLM-generated query decomposition. Experiments on three multi-hop QA benchmarks demonstrate improvements in retrieval coverage and answer accuracy compared with several existing RAG methods, while reducing retrieval latency. The authors also provide ablation studies and additional analyses to investigate the effectiveness of the proposed components.

Critical strengths
  1. The studied problem, reducing the reliance on LLM calls during multi-hop retrieval, is a meaningful problem, particularly for improving the efficiency and scalability of graph-based RAG systems.

  2. The combination of relation-level evidence embeddings and multi-round trajectory search is an interesting approach to multi-hop retrieval without explicit LLM-based query decomposition.

  3. The proposed method achieves competitive or improved performance across three multi-hop QA benchmarks, with promising efficiency gains compared to LLM-guided retrieval approaches.

  4. The paper includes comparisons with multiple baselines, ablation studies, and hyperparameter sensitivity analysis.

Critical weaknesses
  1. Representing relations as vertices and performing multi-hop beam search are closely related to existing graph representations and retrieval techniques. The authors could further clarify which components provide the main methodological novelty beyond combining these existing ideas.

  2. Although the ablation studies demonstrate the importance of relation-level evidence, the current experiments do not fully isolate the contribution of line-graph lifting from relation embeddings and evidence retrieval. A more direct comparison with an edge-based representation using the same evidence embeddings would help clarify this contribution.

  3. The experiments focus on entity-grounded subsets of three benchmarks rather than their complete test sets. Additional evaluation on the full benchmarks, including questions whose answers are not represented as graph entities, would strengthen the generalizability of the results.

  4. While the reported retrieval-time improvements are encouraging, the method still requires LLM-based graph construction and additional relation-level indexing. A more detailed discussion of indexing costs, memory requirements, and scalability to larger or frequently updated knowledge bases would strengthen the practical contribution.

Questions for authors
  1. Could the authors further clarify the key methodological novelty of LiftRAG compared with existing multi-hop retrieval and graph representation methods? Specifically, beyond combining relation-as-vertex mapping with beam search, what are the main conceptual or algorithmic contributions?

  2. To better isolate the benefit of line-graph lifting, have the authors considered comparing it directly against a standard edge-based graph representation using the same evidence embeddings and retrieval settings? How would such a baseline perform compared to the proposed design?

  3. While the authors explain the use of entity-grounded subsets, could they further discuss how LiftRAG would perform on the full benchmark test sets? In particular, how does the framework handle questions whose answers cannot be easily mapped to discrete graph entities?

  4. Given that LiftRAG relies on LLM-based graph construction and relation-level indexing, could the authors provide more details on the additional computational and memory costs of indexing? How well would the approach scale to larger or frequently updated knowledge bases?

1 comment

We would like to share a set of additional experiments that we ran in response to a request from another reviewer, because we believe they are relevant to your assessment as well.

The reviewer asked whether our gains come from the graph mechanism or simply from searching a larger candidate pool. We therefore matched the candidate-pool size in characters across methods and re-ran the comparison, and we also added dense top-100 retrieval followed by a strong cross-encoder reranker under the same 8,000-character budget. At matched pool size our pool-level recall is at parity with dense retrieval and within two points of HippoRAG 2, while our context recall and our conversion from pool to context are the highest on all three datasets. Against dense retrieval with the reranker we lead by 8.2 and 5.1 accuracy points on 2WikiMQA and MuSiQue and tie on HotpotQA, where the dense pool is 1.2 to 2.5 times ours. A coverage analysis further shows that the two kinds of pool reach different evidence, which is the complementarity the paper identifies as future work.

We believe these results further support the novelty and potential of the method, and they may affect your view of it. The full tables (R1 to R4) are in our response to Reviewer Ap7P. We hope this resolves the remaining concerns at once, and we are happy to clarify any number.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.