acceptodds
Under review as a conference paper at ICLR 2027

Rilo: Query Planning with Latent Relational Operators for Multi-Hop RAG

Abstract

Multi-hop retrieval requires connecting evidence across documents and deciding where to search next. The scale and sparsity of the surface-relation space complicate direct planning of multi-hop retrieval. We introduce RILO, a retrieval planner that learns a compact vocabulary of composable latent relational operators. These operators capture transformations shared across surface relations and are trained on corpus-derived paths to predict evidence endpoints. Given a question and retrieved evidence, RILO composes the operators into candidate multi-step plans. Each plan and its predicted endpoint guide a single corpus query without online graph traversal. A budget-aware controller selects the next query or stops based on expected evidence gain. RILO then updates its evidence state and replans as new passages arrive. In controlled evaluations on 2WikiMultiHopQA, HotpotQA, and MuSiQue, RILO achieves 73.10% macro complete-support coverage within the top ten passages (SC@10). With at most three corpus calls, it exceeds the strongest evaluated baseline by 3.86 percentage points.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-hop Rag, knowlege Graph, Query Plan

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8nhTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Rilo, which is a method for multi-hop retrieval. It learns a set of relation operators and combines them to predict what evidence should be searched next. These predictions are used to generate search queries. A controller decides which query to use or when to stop searching based on the expected benefit and remaining search budget.

Critical strengths
  • The paper proposes a method that uses learned relation operators to plan the next search. It combines multiple relation steps into a single search query and updates the plan when new documents are retrieved.

  • The paper includes matched controls, ablation studies, experiments with larger document collections, and cross-dataset transfer. It also evaluates the learned operators on relation combinations not seen during training, which helps assess their ability to generalize.

Critical weaknesses
  • The proposed method does not show a clear overall improvement over Direct-continuous in the 96-token setting, and its lower performance on MuSiQue is statistically significant. In the 32-token setting, the improvement on MuSiQue is also not statistically significant. These results raise questions about when the learned operators are actually useful.
  • The improvements in answer F1 are also small, and RILO performs worse than EfficientRAG on MuSiQue. Given the number of models and the complexity of the proposed method, the additional benefits over simpler approaches are not sufficiently demonstrated.
Questions for authors
  • What explains the different results against Direct-continuous across the two query-length settings, particularly the reversal on MuSiQue? Is there evidence that separates errors in operator composition from errors in query generation?
  • The short-query protocol specifies a 32-token output limit for all matched methods, but Direct-matched and Direct-continuous report the same scores as in the 96-token comparison (Tables 9, 10, 16). Were these results obtained from separate evaluations? Please clarify why the scores are identical.
Reviewer_kUX1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

RILO learns a small vocabulary of discrete "latent relational operators" from relation paths in an LLM-extracted corpus graph. Starting from entities in the question and the retrieved evidence, it composes these operators into multi-step plans, turns each plan into one text query with a small T5, and uses a budget-conditioned Q-function to choose the next query or stop. On 2Wiki, HotpotQA and MuSiQue, it improves complete-support retrieval over the strongest baseline, with a marginal gain in answer F1.

Critical strengths
  1. Consistent retrieval gains under a controlled comparison. Retrieval coverage improves on all three datasets with a shared index, reader and call budget, and the paper includes supervision-matched controls.
  2. A sensible planning idea. Issuing one query per multi-step plan separates lookahead depth from the number of corpus calls, and deeper plans help at a fixed budget (Figure 5b).
Critical weaknesses
  1. Hard to read. The prose consists of short, disconnected sentences and many coined terms ("evidence-anchored plans", "endpoint-distinct continuations").
  2. Several modules are under-specified. The transition model GG and the T5 query realizer are specified concretely (Appendices A.2 and A.7.2). The router FF, the proposal pηp_\eta, the value network QQ and the direct query generator CdirectC_{\text{direct}}, however, are described only by their inputs (and, for some, parameter counts). The paper does not say what these networks are, or how the question, passage history and action history are encoded as their inputs.
  3. Limited novelty. Most components are known: straight-through code routing, recurrent path embeddings, HyDE-style query generation, and fitted-Q stopping.
  4. The core idea adds little over strong controls. In the configuration used for the main results, RILO is statistically indistinguishable from the continuous-state control (Direct-continuous) and worse on MuSiQue (Tables 9 and 12). The learned operators are also not semantically coherent (Table 1).
Questions for authors
  1. What are the architectures of FF, pηp_\eta, QQ and CdirectC_{\text{direct}}, and how is the retrieval state sts_t (question, passage history, action history) encoded as their input?
  2. Given Tables 9 and 12, what supports the claim that discrete operators, rather than a continuous state, drive the gains? Are the Direct-matched/Direct-continuous scores in Tables 9 and 16 from the same runs?
  3. How does the preparation cost compare with HippoRAG2 and IRCoT?
Reviewer_5E4oOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RILO, a multi-hop RAG retrieval planner that learns a compact vocabulary of latent relational operators from corpus-derived relation paths through endpoint-prediction supervision. At inference time, operators are composed into candidate multi-step continuations anchored in observed entities, and each plan is realized as a single endpoint-guided text query without online graph traversal. A budget-conditioned value controller selects retrieval actions or stops, and replans as new evidence arrives. Across 2WikiMultiHopQA, HotpotQA, and MuSiQue, RILO achieves 73.10% macro complete-support coverage (SC@10) within at most three corpus calls, improving on the strongest evaluated baseline by 3.86 percentage points. Overall, the method offers an interesting connection between learned relational abstraction and budgeted retrieval, although the independent contribution of discrete operators merits more careful qualification.

Critical strengths

S1. The paper addresses an important multi-hop RAG problem: anticipating where missing evidence may lie within a limited corpus-access budget, rather than relying on locally relevant retrieval or online graph-path expansion.

S2. The central idea is coherent and memorable. A compact compositional relation vocabulary supports multi-step relational lookahead, while an endpoint-conditioned query realizes a proposed continuation through one corpus search.

Critical weaknesses

W1. Insufficient clarification of the core methodological contributions and evaluation protocols. W2. Limited discussion of practical robustness and computational cost.

Questions for authors

D1. Under MATCH-96, RILO shows only marginal improvement over Direct-continuous, whereas the gain is more pronounced under MATCH-32 (Appendices B.1 and D). Could the authors explain this discrepancy and further clarify the independent contribution of discrete relational operators?

D2. The MAIN, MATCH-96, and MATCH-32 protocols report different results (Appendix Table 19), while Figure 5(a) adopts the short-query setting. The authors are encouraged to further clarify the differences among these evaluation protocols.

D3. The related work could further discuss prior studies on compositional reasoning in knowledge graph embedding spaces, such as Guu et al. (2015), GQE (2018), and Query2Box (2020). These methods share similar ideas of composing latent relational transformations to predict multi-hop targets. The authors are encouraged to clarify how RILO differs from these approaches.

D4. RILO relies on extracted graph relations and observed pivot entities for planning (Appendix A.1; Section 4.2). How would entity-resolution errors or missing key bridge entities affect subsequent planning, and how does the current framework handle such cases?

D5. Although RILO realizes multi-step relational plans through a single corpus search, its retrieval latency remains higher than that of some baselines (Appendix E.3), raising concerns about its overall computational cost.

D6. Minor presentation issues: Figure 1 contains a typo (“writen_by” should be “written_by”). In addition, Figure 2 highlights K=8 as the selected setting, whereas the main experiments use K=16. The authors are encouraged to clarify this distinction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.