acceptodds
Under review as a conference paper at ICLR 2027

SPEAR: Privacy- and Executability-Aware Subquery Routing for Edge-Cloud Collaborative RAG

Abstract

Retrieval-augmented generation (RAG) in financial, medical, and enterprise applications often depends on private local data, creating a structural mismatch: edge devices can access these data but have limited reasoning capabilities, while powerful cloud models cannot directly access them. Existing edge-cloud methods largely provide coarse-grained sketches or guidance, limiting their effectiveness on multi-step queries involving mixed public-private knowledge. We propose SPEAR, a subquery-level, privacy- and executability-aware routing method for edge-cloud collaborative RAG. SPEAR desensitizes the user query for cloud-side decomposition into a dependency-structured subquery graph, then jointly estimates each subquery's privacy dependency and edge executability to select its execution site. Subqueries are executed collaboratively according to graph dependencies, and their outputs are fused at the edge. Experiments across multiple privacy-sensitive domain datasets show that SPEAR consistently improves generation quality while limiting private knowledge exposure. Code is available at https://anonymous.4open.science/r/SPEAR-5D6D.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-Augmented Generation, Edge–Cloud Collaboration, Privacy-Preserving Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KApNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new method to address the privacy-efficiency trade-off in the context of cloud-edge computing platform. Essentially, edge devices are usually considered trustworthy to touch on the private data but cannot deliver powerful computation; therefore the computation-heavy portion is offloaded to the cloud while sensitive data should be protected. The paper is fairly complete: I see the proposed method is extensively evaluated against multiple datasets and baselines.

Critical strengths
  • Evaluation is pretty substantive
  • The problem statement is also clear
Critical weaknesses
  • Novelty is somewhat limited. Almost each of the components in SPEAR has a close counterpart in the literature. It's unclear what the technical challenge is to combine those existing components

  • The method lacks a systematic design. While the method itself is clear, but I wonder why it was designed like this. To me, the design seems highly heuristic without much intuition behind. If the design was, for example, derived from experiments, the authors should have clarified that.

Questions for authors
  • Please articulate the novelty of this paper

  • Please discuss more about the design space.

1 comment

Response to the Novelty, Design Rationale, and Technical Challenges of SPEAR.

Thank you for raising these important concerns. The current manuscript presents PDE, EEE, and ECR primarily as finalized components, which may obscure the observation-to-design process and give the impression that SPEAR simply combines existing techniques. However, these components not introduced independently or simply assembled together. Our design began with a diagnostic analysis of decomposed subqueries, which revealed two distinct placement failures: privacy-critical subqueries can be incorrectly offloaded to the cloud, whereas subqueries beyond the reliable capability of the edge SLM can be incorrectly retained locally. These two risks cannot be addressed independently because both privacy dependency and edge executability are conditioned on the parent context in the SQ-DAG.

To quantify these observations, we evaluate two simple placement rules. Privacy-to-Cloud Misrouting is the proportion of cloud-routed subqueries that are independently annotated as requiring local execution for privacy reasons. Edge-Execution Failure Rate is the proportion of edge-routed subqueries whose locally generated answers fail an independent correctness-and-sufficiency evaluation.

DatasetPrivacy-to-Cloud Misrouting under Executability-Driven PlacementEdge-Execution Failure under Privacy-Driven Placement
P-PubMedQA43.7%36.8%
P-FinQA33.8%26.9%
P-MultiHopRAG58.4%42.5%

The diagnostic results reveal three recurring patterns:

  1. Executability alone is insufficient. An executability-driven rule offloads subqueries that are difficult for the edge SLM, but some of these subqueries still require private evidence. This results in substantial privacy-to-cloud misrouting—33.8% to 58.4% across the three datasets. This observation motivates PDE, which estimates private dependence through both counterfactual private-aware gain and structural private-slot necessity.

  2. Privacy dependence alone is insufficient. A privacy-driven rule retains privacy-critical subqueries locally, but privacy dependence does not imply that the edge SLM can resolve them reliably. Consequently, 26.9% to 42.5% of locally executed subqueries fail the independent correctness-and-sufficiency evaluation. This observation motivates EEE, which estimates local execution reliability using both answer quality and conclusion-level self-consistency.

  3. The two risks are dependency-conditioned. Privacy dependency and edge executability cannot be treated as independent, static attributes. Both are evaluated under the parent context hih_i, because inherited answers may change the private information required by a descendant subquery as well as the ability of the edge SLM to resolve it. This motivates ECR, which applies the privacy-prioritized hard rule when both PDi≥τp\mathrm{PD}_i \ge \tau_p and EEi≥τe\mathrm{EE}_i \ge \tau_e, and uses the joint soft score for the remaining ambiguous cases.

Thus, SPEAR’s contribution does not lie in claiming that decomposition, privacy assessment, or capability assessment is individually new. Rather, it lies in formulating private edge-cloud RAG as a dependency-aware subquery placement problem under two competing and asymmetric risks, and in deriving a coupled estimation and routing mechanism to address that problem.

Reviewer_K5icOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents SPEAR, a privacy- and executability-aware subquery routing framework for edge-cloud collaborative retrieval-augmented generation (RAG). The paper addresses the challenge of answering complex queries involving both public and private knowledge, where edge-side small language models have access to sensitive information but limited reasoning capabilities, while cloud-side large language models provide stronger reasoning capabilities but cannot directly access private data.

SPEAR first desensitizes the original query and delegates its decomposition to a cloud-side LLM, producing a dependency-structured subquery graph (SQ-DAG). It then introduces a Privacy Dependency Estimator (PDE) to quantify each subquery's reliance on private knowledge and an Edge Executability Estimator (EEE) to evaluate whether the edge-side model can reliably execute it. Based on these estimates, an Edge-Cloud Routing (ECR) module assigns subqueries to the edge or cloud using a combination of hard rules and threshold-based scores. The resulting subquery answers are finally integrated at the edge.

Critical strengths
  1. A practical and real-world problem.

The paper investigates an important challenge in privacy-sensitive edge-cloud collaborative RAG: when complex/multi-turn queries require reasoning over both public and private knowledge. The mismatch between the computational capabilities of edge devices and the privacy constraints of cloud models is a practical problem that should be addressed.

Critical weaknesses
  1. The proposed design is insufficiently justified.

The proposed methodology appears to be a straightforward integration of existing query decomposition and heuristic routing techniques. The central assumption that decomposing a complex query into an SQ-DAG improves multi-step reasoning is not sufficiently justified. Decomposition does not inherently reduce reasoning complexity and may introduce additional challenges, such as information fragmentation, incorrect dependency identification, and error propagation across intermediate subqueries. The observed improvements could instead arise from increased cloud-model participation or additional inference steps rather than the proposed graph-based routing mechanism.

  1. The baseline comparison is insufficient and existing multi-step and multi-hop reasoning methods are not compared.

The paper compares SPEAR against several edge-cloud collaborative generation methods, including RemoteRAG, MINIONS, CoGenesis, LSRP, and PRISM. Although these baselines are relevant to privacy-preserving collaborative inference, most do not explicitly incorporate dependency-aware multi-step query decomposition and execution. For example, the literature mentioned below is somewhat relevant with multi-step RAG-based reasoning and should not be neglected.
Optimizing Question Semantic Space for Dynamic Retrieval-Augmented Multi-hop Question Answering (ACL, 2025) Question Decomposition for Retrieval-Augmented Generation (ACL SRW 2025) Planned Active Retrieval and Reasoning for Multi-Hop Question Answering (arXiv, March, 2026)

Questions for authors
  1. Methodological Novelty and Justification. What fundamentally distinguishes SPEAR's SQ-DAG-based decomposition and heuristic routing strategy from existing multi-step reasoning and query decomposition approaches? More importantly, why should decomposing complex queries into dependency-structured subqueries improve reasoning quality, particularly given the potential for dependency errors and error propagation?

  2. Baseline Fairness and Effectiveness Attribution. How does SPEAR compare against recent multi-hop reasoning approaches, such as Q-DREAM (ACL 2025), when adapted to the same privacy-aware edge-cloud setting? Can the authors demonstrate that the performance improvements originate from the proposed privacy- and executability-aware routing mechanism rather than conventional query decomposition, increased cloud-model participation, or additional inference budgets?

1 comment

Response to the proposed design is insufficiently justified.

Thank you for raising this important concern. We agree that query decomposition does not inherently reduce reasoning complexity and may introduce information fragmentation, incorrect dependencies, and error propagation. Our claim is more specific: the SQ-DAG is not intended as a generic mechanism for improving reasoning. Instead, it serves as a control representation that makes inter-subquery dependencies explicit and enables fine-grained execution placement. Because a subquery’s privacy dependency and edge executability may change according to its parent outputs, independently routing each subquery could either expose private-derived context to the cloud or propagate unreliable edge-generated results.

The contribution of dependency-aware routing is supported by the controlled ablations in Sec. 4.4 (Tables 3 and 6) and Appendix F.4 (Table 22). All variants operate on the same SQ-DAG and follow the same multi-step pipeline, while only the routing components or decision rules are changed. Removing ECR, its hard privacy-first rule, or its soft trade-off rule consistently degrades both answer quality and privacy protection. These results demonstrate that the execution-placement policy contributes materially even when the decomposition and reasoning steps remain unchanged.

To further examine whether SPEAR’s gains merely result from a larger inference budget, we compare all methods within matched total-output-token intervals:

Answer QualityPrivacy Leakage
Output-budget rangeMethodF1 ↑SM ↑LaJ ↑AUC ↓MI ↓
[7.0K,7.5K)RemoteRAG.534.691.688.683.225
CoGenesis.500.598.575.607.193
LSRP.478.592.560.582.184
PRISM.517.659.623.550.180
SPEAR.558.731.699.548.178
[7.5K,8.0K)RemoteRAG.539.698.694.683.225
CoGenesis.506.604.581.609.195
LSRP.486.602.569.584.186
PRISM.524.667.632.551.181
SPEAR.566.741.707.548.178
[8.0K,8.5K]RemoteRAG.543.704.699.684.226
CoGenesis.512.611.587.612.197
LSRP.494.612.578.588.188
PRISM.531.676.640.553.183
SPEAR.573.750.714.547.177

SPEAR consistently achieves the best answer quality and privacy–utility balance within every output-budget interval. Increasing the budget yields only moderate improvements for all methods, while low-budget SPEAR remains stronger than higher-budget alternatives.

Response to the Coverage of Multi-Step and Multi-Hop Reasoning Baselines

Thank you for this important suggestion. Our original baseline selection was guided by SPEAR’s privacy threat model. Conventional multi-hop reasoning methods are designed for centralized execution and provide no mechanism for protecting private queries, retrieved private evidence, or intermediate reasoning states. Applying them directly would expose the complete reasoning process to the cloud. We therefore initially focused on privacy-preserving and edge-cloud collaborative baselines.

Nevertheless, we agree that these methods provide valuable quality-oriented, non-private reference points for assessing how much multi-hop reasoning capability SPEAR preserves under privacy constraints. We therefore additionally evaluate QD-RAG, Q-DREAM, and PAR²-RAG on P-MultiHopRAG using the same cloud LLM, retriever, and inference-budget limit.

MethodF1 ↑SM ↑LaJ ↑AUC ↓MI ↓Privacy Protection
Cloud-Only.604.787.743N/AN/ANone
QD-RAG.603.779.738N/AN/ANone
Q-DREAM.608.793.744N/AN/ANone
PAR²-RAG.611.795.752N/AN/ANone
SPEAR (Ours).587.767.724.546.175Yes

As expected, unrestricted multi-hop methods achieve moderately higher answer quality because the cloud can directly access all queries, evidence, and intermediate reasoning states. Nevertheless, SPEAR remains within 2.8% of the strongest multi-hop baseline across F1, SM, and LaJ, while satisfying the private edge-cloud threat model and maintaining a privacy-inference AUC close to random guessing.

The results show that SPEAR preserves most of the reasoning capability of specialized multi-hop methods while additionally supporting privacy- and executability-aware placement.

Reviewer_cUzkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the capability and privacy mismatch in edge-cloud collaborative RAG, where edge devices have private data but limited compute, while powerful cloud models face privacy constraints. To resolve this, the authors propose SPEAR, a fine-grained routing framework that decomposes desensitized queries into a subquery dependency graph, dynamically assigning each node to the edge or cloud based on joint estimations of privacy dependency and edge executability. Experimental results show that SPEAR consistently improves generation quality closer to the cloud upper bound while rigorously minimizing private knowledge exposure across all transmission stages.

Critical strengths

The conceptual structure of the proposed approach is clearly illustrated and well-documented through high-quality figures and comprehensive text.

The paper provides a relatively comprehensive mathematical formulation and theoretical analysis.

Critical weaknesses

There are glaring flaws in the theoretical derivations. E.g., Appendices B.4 and B.5 specify different values for N yet exhibit identical subsequent derivation contents.

The experimental design exhibits critical flaws in its logic. E.g., the ablation results in Table 3 and Table 19 present an unexplained paradox where the routing algorithm (ECR) somehow operates and yields specific metrics (AUC=0.898 and MI=0.459) even when both the PDE and EEE are disabled. Given that the routing decisions in Eq. (13) strictly rely on scores from PDE and EEE, it remains completely opaque what signals the ECR utilizes to route subqueries in their absence.

Questions for authors

Did the main experiments actually incorporate DP mechanism? If so, why were the necessary specification omitted from the experimental setup? If not, what is the justification for dedicating substantial space to the theoretical DP guarantees?

2 comments

Response to the (N)-Setting Inconsistency in Appendices B.4–B.5.

Thank you for pointing this out. We acknowledge that the inconsistency between Appendices B.4 and B.5 is an editing error in the illustrative budget configuration. We will remove the redundant presentation and provide a single, internally consistent configuration in the revised manuscript. This issue is confined to the numerical illustration; the general composition result and its proof remain unaffected. We apologize for the oversight.

Response to the ECR Routing Mechanism under Joint PDE–EEE Ablation.

Thank you for raising this question. The “ECR-only” setting does not allow ECR to infer routes from an unreported signal. In the ECR-only ablation, ECR does not infer routes from an alternative signal. The missing PDE and EEE outputs are replaced by fixed zero priors, i.e., PDi=EEi=0\mathrm{PD}_i=\mathrm{EE}_i=0. Consequently, neither routing condition in Eq. (13) is satisfied, and all subqueries follow the predefined cloud-fallback branch. This is a deliberately degenerate control rather than a fully functional ECR configuration. Its high leakage is expected because privacy-dependent subqueries are indiscriminately offloaded, while its low answer quality results from executing desensitized and decomposed subqueries without access to the required private evidence. We will explicitly state this protocol and relabel the row as “ECR with fixed-zero inputs (cloud fallback).”

Response to the Role of DP in the Main Experiments and Theoretical Analysis.

The main experiments use empirical SPEAR without DP noise, while the DP analysis characterizes a separate certified execution mode. We include it because empirical AUC and leakage estimates are attacker-dependent, whereas the DP result specifies how SPEAR can obtain an attacker-independent guarantee over the complete cloud-visible transcript. We agree that this distinction and the current allocation of space are unclear; we will state the two modes explicitly, retain only the principal guarantee in the main text, and move the detailed derivations to the appendix.

  1. “(N)-Setting Inconsistency in Appendices B.4–B.5.” is just one example of “glaring flaws in the theoretical derivations” (like lack of the support consistency guarantee of the exponential mechanism in DP); “ECR Routing Mechanism under Joint PDE–EEE Ablation” is just one example of “critical flaws in its logic” in the experimental design (like lack of token-budget-matched comparisons). They are not exhaustive.
  2. As for “ECR Routing Mechanism under Joint PDE–EEE Ablation”, the authors replied that “all subqueries follow the predefined cloud-fallback branch”, “privacy-dependent subqueries are indiscriminately offloaded”, and “cloud fallback”. Since all subqueries fall back to the cloud and the authors have defined “the cloud is honest-but-curious” in 2.2, then the values AUC=0.898 and MI=0.459 are unreasonable and lacking sufficient explanation.

As substantial flaws remain in the core methodology and theoretical derivations, the authors’ comment is insufficient to warrant raising the rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.