acceptodds
Under review as a conference paper at ICLR 2027

Spend Where It Matters: Fine-Grained Privacy Allocation for Edge-Cloud Reasoning

Abstract

Privacy-sensitive multi-step reasoning must exploit the capabilities of cloud large language models (LLMs) without exposing private knowledge stored on trusted edge devices. Local-only execution protects privacy but limits reasoning quality, whereas existing edge-cloud approaches typically make coarse request-level decisions that cannot accommodate the heterogeneous privacy and quality requirements of individual reasoning steps. We formulate this problem as fine-grained privacy allocation over a reasoning graph and introduce Operational Privacy Dependency to characterize step-specific privacy-utility frontiers. We further derive a lower bound on the privacy cost required to achieve a target reasoning distortion and establish an adaptive end-to-end privacy guarantee. Building on this theory, we propose Swim, which constructs privacy-valid actions, analytically estimates their state-dependent quality-privacy operating points, and performs frontier-aware online allocation under a global privacy budget. Across six dataset-model settings, Swim retains 95.7% of Cloud-Only F1, improves F1 by 5.6% over the strongest non-Cloud-Only baseline, and reduces privacy-leakage AUC and MI by 18.3% and 39.1%, respectively. Code is available at https://anonymous.4open.science/r/Swim-4EB3/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Edge–Cloud Collaborative Inference, Privacy–Utility Trade-off

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

The paper looks at privacy in edge-cloud reasoning, where a small local model holds private knowledge and a cloud LLM helps with a multi-step reasoning graph. Instead of one privacy decision per request, the authors split a global DP budget across the individual steps. Their method, Swim, scores each step with local heuristics (uncertainty, importance, dependence on private evidence) and picks edge execution or a cloud call with a certain noise level. The authors also define “Operational Privacy Dependency”, prove a lower bound, and cite adaptive composition. They test on three QA datasets they built themselves and report better quality and lower attack scores than other edge-cloud methods.

Critical strengths
  1. The question is good. Where a limited privacy budget should be spent in multi-step reasoning is a real and under-studied problem.
  2. Counting the incremental privacy cost with an accountant and enforcing a hard budget is a sensible design.
Critical weaknesses
  1. My main concern is that the privacy claim seems to differ from the system in the experiments. The paper says that choosing actions costs no privacy because everything is computed locally. But the scores that decide whether a step goes to the cloud, and with how much noise, are computed from private evidence. Whether the cloud is called at all is therefore visible to the cloud and depends on private data. The paper’s own appendix says such routing decisions must be counted when they depend on protected information, but the method does not do this. The same goes for the threshold that allows “privacy-free” cloud calls. That threshold is effectively an unaccounted privacy knob, and the authors’ own sensitivity analysis shows leakage rising when it is increased. The theorem only holds if each release is DP given the history, and nothing shows this holds for the real pipeline. So “better quality under the same formal privacy budget” is not established.

  2. My second concern is that it is unclear how much privacy is actually being spent. The paper does not state the noise mechanism, sensitivity or noise scales. With the listed settings (64-dim state, clip norm 1, ε at most 1), the usual Gaussian mechanism would give noise much larger than the signal. The leakage numbers also hardly change when the protected releases are added to the transcript. This suggests the quality gains may mostly come from the free cloud calls and not from careful spending of privacy. If so, the main story of the paper is not supported.

Overall, the paper is hard to follow in terms of the mathematical content. The conceptual content is clearer, but it seems that the mathematical notation and explanation could be significantly improved and clarified.

Questions for authors
  1. What exactly is sent in a “privacy-free” action, and is it independent of unreleased private evidence? How are routing decisions and the threshold gating counted in the guarantee?
  2. What are the DP mechanism, sensitivity and noise scales? How much of the quality gain remains if only privacy-free actions are used?
  3. Can you compare against request-level, equal-split and greedy allocation using the same mechanism and accountant?
  4. Does the predicted gain actually rank real per-action gains correctly on held-out data?
  5. In the appendix you describe the "sanitization" procedure. However, there is not guarantee that an LLM will actually remove all personal and sensitive information. Can you please comment on this?
3 comments

Response to privacy accounting for routing, threshold gating, and “privacy-free” actions.

Thank you—this concern reveals an ambiguity in our terminology. An edge action sends no request to the cloud. An r0r_0 action sends only a cloud-safe payload containing the public node description and predecessor outputs already present in the cloud-visible history. Raw queries, private evidence, private slots, and unreleased predecessor outputs are excluded. Therefore, r0r_0 has zero private-payload cost.

We agree that the invocation pattern may itself depend on private evidence. The cloud observes whether it is called and which tier is selected, while these decisions depend on the locally computed private-dependency score and threshold. In the reported implementation, the formal accountant composes the protected payload releases rℓr_\ell for ℓ≥1\ell\geq1, but does not assign a separate ϵ\epsilon to deterministic routing or threshold gating. These observable routing events are included in the metadata-only and full cloud-visible transcripts used in our empirical leakage evaluation.

Response to the DP configuration, privacy expenditure, and sources of utility gains.

Thank you for identifying the missing mechanism details. We agree that the original protected releases operate in a high-noise regime. Each of at most m=16m=16 release units is clipped to C=1C=1 and averaged using a fixed denominator, giving replace-one sensitivity Δ2=2C/m=0.125\Delta_2=2C/m=0.125. We then apply a Gaussian mechanism calibrated through zCDP with atomic δ=10−6\delta=10^{-6}. The per-query privacy spend is obtained by composing the selected protected actions. We will add the sensitivity, calibration, noise scales, and composed privacy spend to the paper.

Table R1 confirms that, under the original configuration, the RMS noise norm of 5.355.35–26.3826.38 exceeds the clean-state norm upper bound of 11. However, differential privacy limits distinguishability rather than making the release independent of the private state. For a release s~=s+z\tilde{s}=s+z with zero-mean Gaussian noise, E[s~∣s]=s\mathbb{E}[\tilde{s}\mid s]=s: the noise increases variance but does not remove the signal in expectation. Moreover, the RMS norm aggregates noise over all 64 dimensions, whereas task utility may depend on a smaller number of task-relevant projections. The cloud model can combine this residual information with the public node description and cloud-visible context. Consequently, the protected actions achieve actual gains of 0.2140.214–0.2390.239, compared with 0.2050.205 for r0r_0.

We also evaluate lower-noise settings. As the RMS noise decreases to 2.722.72 and 1.401.40, the maximum actual gain increases to 0.2930.293 and 0.3000.300, respectively. This confirms that reducing the perturbation improves the utility of protected releases, while the positive gains under the original configuration show that useful protected information remains recoverable despite the large noise.

SettingActionϵ\epsilonRMS noiseActual gain
OriginalFree cloud r0r_0000.205
OriginalProtected r1r_10.226.380.214
OriginalProtected r2r_20.413.240.215
OriginalProtected r3r_30.68.850.219
OriginalProtected r4r_40.86.660.224
OriginalProtected r5r_51.05.350.239
2×ϵ2\times\epsilonProtected r1r_10.413.240.215
2×ϵ2\times\epsilonProtected r2r_20.86.660.224
2×ϵ2\times\epsilonProtected r3r_31.24.470.256
2×ϵ2\times\epsilonProtected r4r_41.63.380.282
2×ϵ2\times\epsilonProtected r5r_52.02.720.293
4×ϵ4\times\epsilonProtected r1r_10.86.660.224
4×ϵ4\times\epsilonProtected r2r_21.63.380.282
4×ϵ4\times\epsilonProtected r3r_32.42.280.297
4×ϵ4\times\epsilonProtected r4r_43.21.730.299
4×ϵ4\times\epsilonProtected r5r_54.01.400.300

Response to the clarity of the mathematical notation and exposition.

Thank you for this helpful feedback. We agree that the mathematical presentation can be improved. We will revise the paper to define all symbols before use, standardize notation across sections, clarify the assumptions and intuition behind the key equations, and separate the mechanism, action-selection rule, and privacy accounting more clearly. We will also add a concise overview to make the mathematical development easier to follow.

Response to the DP specification, utility of privacy-free Actions, and accuracy of per-action gain prediction.

Thank you for requesting these clarifications. For each protected action, we encode at most m=16m=16 private evidence units into 64-dimensional signed-hash vectors. Each vector is clipped to L2L_2 norm C=1C=1, and the aggregate uses the fixed denominator mm:

\qquad n\le m.$$ Under replace-one adjacency, $\Delta_2=2C/m=0.125$. We release $$\widetilde{s}=s(D)+\mathcal N(0,\sigma^2I_{64}).$$ With per-release $\delta=10^{-6}$, we calibrate the Gaussian mechanism through zCDP: $$\rho=\left(\sqrt{\log(1/\delta)+\epsilon} -\sqrt{\log(1/\delta)}\right)^2, \qquad \sigma^2=\frac{\Delta_2^2}{2\rho}.$$ For adaptive releases, the accountant composes the corresponding $\rho$ values and converts the total back to an $(\epsilon,\delta)$ guarantee. We further evaluate all actions from matched held-out pre-action states. Actual quality denotes the realized downstream quality after taking action $a$, while $\widehat g(a)$ is the gain predicted by the action selector. | Action | $\epsilon$ | Noise scale $\sigma$ | Predicted gain $\widehat g(a)$ | Actual quality | | ------------------ | ---------- | -------------------- | ------------------------------ | -------------- | | Privacy-free $r_0$ | 0 | — | 0.157 | 0.205 | | DP release | 0.2 | 3.297 | 0.166 | 0.214 | | DP release | 0.4 | 1.655 | 0.168 | 0.215 | | DP release | 0.6 | 1.107 | 0.169 | 0.219 | | DP release | 0.8 | 0.833 | 0.171 | 0.224 | | DP release | 1.0 | 0.669 | 0.175 | 0.239 | **Response to comparisons with alternative privacy-budget allocation strategies.** Thank you for this suggestion. We added three allocation baselines while keeping the DP mechanism, action space, Gaussian noise calibration, zCDP accountant, global $(\epsilon,\delta)$ budget, models, and test examples identical: \- Request-level: selects one privacy tier for the entire request; \- Equal-split: divides the global privacy budget uniformly across private candidate nodes; \- Greedy: selects the action with the largest immediate predicted gain without reserving budget for future nodes. | Allocation policy | F1 ↑ | SM ↑ | LaJ ↑ | AUC ↓ | MI ↓ | | ----------------- | ---- | ---- | ----- | ----- | ---- | | Request-level | .647 | .663 | .669 | .566 | .184 | | Equal-split | .661 | .678 | .695 | .574 | .187 | | Greedy | .672 | .681 | .723 | .607 | .193 | | Swim | .691 | .729 | .783 | .523 | .158 | Under the same formal privacy budget, Swim achieves both the highest answer quality and the lowest empirical leakage. Compared with Greedy, the strongest allocation baseline in answer quality, Swim improves F1, SM, and LaJ by 1.9%, 4.8%, and 6.0%, respectively, while reducing AUC and MI by $.084$ and $.035$. These results isolate the benefit of Swim's dependency-aware allocation: request-level and equal-split policies cannot exploit heterogeneous node-level privacy–utility returns, whereas Greedy can exhaust the budget on early nodes. Swim instead reserves and allocates privacy budget according to both the current return and downstream reasoning dependencies. **Response to the privacy guarantees of LLM-based sanitization.** Thank you for the question. Swim performs sanitization locally by replacing every field labeled private with a typed placeholder while retaining the task structure required for cloud-side decomposition. Before transmission, the sanitized representation is validated against a fixed schema that admits public task metadata, reasoning operators, dependency relations, and typed placeholders. We will add the exact sanitization prompt, the admissible output schema, and this formal statement to the appendix.
Reviewer_bVKpOpenReview ↗

Response to Authors

Thank you for the detailed response, I appreciate the effort the authors put into their rebuttal, but I fear my main concerns are confirmed rather than resolved, especially in the context of other reviewer's comments.

1. Privacy claim. You agree that routing and threshold gating depend on private evidence and are not charged to the accountant. Including them in an empirical leakage evaluation is not a formal guarantee, so "better quality under the same formal privacy budget" does not currently hold. Either account for these decisions in the guarantee or restate the claim as empirical only. The same applies to sanitization. It is described as replacing fields "labeled private," but real queries are unlabeled, so its recall remains untested, and the DP guarantee does not cover the query itself.

2. Where the gains come from. As far as I can tell, your tables show that a privacy-free cloud call already gives an actual gain of 0.205, and protected releases at the original settings add only about 0.01 to 0.03. Cutting the noise roughly 3x (ε from 0.2 to 0.6) barely changes the gain. Large gains appear only at per-release ε far above the global budget. The benefit therefore seems to come mostly from free calls, not from spending privacy.

3. Scope. I think the paper tries to do too much at once. In one paper you have a theoretical framework, a multi-component heuristic allocator, an unspecified sanitizer, three new datasets and a mobile demo, all evaluated end to end. This makes it really hard to follow and also each step requires sets of experiments and ablations etc. which cannot be effectively dealt with in one paper. I would advise the authors to maybe to split off some core ideas and write their paper around that, rather than doing too much. There are some nice ideas and implementations in this paper, but in my opinion there is too much going on. Other reviewers may disagree with this point, so it won't be my deciding factor, but this is just a point I wanted to note.

4. Presentation. Finally, I share some of the concerns about notation and exposition raised by Reviewer 7P2s. There is a lot of unclear notation with unnecessary amount of complexity. The problems seem pervasive and tied to the technical content, so I do not think they can be fixed in one revision round.

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

This paper studies edge-cloud collaborative multi-step reasoning in privacy-sensitive scenarios, aiming to leverage the reasoning capabilities of large language models in the cloud while protecting the private knowledge stored on trusted edge devices. Addressing the difficulty of request-level collaboration in adapting to the privacy and quality requirements of different reasoning steps, the authors model the problem as fine-grained privacy budget allocation on a reasoning graph, propose operational privacy dependence, derive privacy-utility lower bounds, and conduct adaptive end-to-end privacy composition analysis. On this basis, the Swim framework is proposed, which constructs privacy-efficient candidate actions and analytically estimates state-related quality-privacy trade-offs to perform frontier-aware online allocation under a global privacy budget. The paper reports that, across six dataset-model settings, Swim retains 95.7% of the F1 score of the Cloud-Only model, represents a 5.6% relative improvement over the strongest non-Cloud-Only baseline in F1, and reduces empirical privacy leakage indicators AUC and MI by 18.3% and 39.1%, respectively.

Critical strengths

1.The research problem is practically significant, and the problem modeling is clear. The paper focuses on edge-cloud collaborative inference in privacy-sensitive scenarios, incorporating the differences in private information and cloud capabilities at different inference steps into a unified budget allocation problem. Fine-grained allocation based on the inference graph provides a valuable perspective for studying the privacy-utility trade-off in multi-step tasks.

2.The method design is coherent and can balance current benefits with future requirements. Swim estimates candidate action work points by combining local uncertainty, task importance, and the dependence on private information, and then selects actions through dynamic privacy pricing and hard budget constraints. Reserving budgets for downstream nodes is a reasonable design, and local analytical estimation avoids repeatedly running cloud candidates to evaluate benefits.

  1. Future budget pricing may overestimate downstream privacy demand. Equations (7) and (9) use unresolved descendants’ semantic importance to determine the current privacy price, without distinguishing steps requiring new private releases from those using public or already released information. This may encourage overly conservative budget reservation. Reporting budget utilization in cases with many privacy-free descendants would help assess this concern.
Critical weaknesses

1 Equation (6) in §4.3 is crucial for action selection, and its form draws inspiration from the signal-to-noise ratio recovery factor in a signal-plus-Gaussian-noise model. However, U_i*D_i^sem has not been proven to be a valid proxy for "signal strength" or "recoverable private information volume." Although intuitive and heuristic as surrogates, the authors should ideally provide experimental or theoretical evidence to demonstrate the superiority and effectiveness of this formulation.

2 §5.5 only examines the global privacy budget and dependency threshold, without evaluating the validity and rationality of the local sampling number M=5 and kernel bandwidth h=1. The former affects estimation stability and local computational cost, while the latter directly influences dependency assessment and gating outcomes. Supplementing with experiments on a few representative parameter settings would help clarify the method's reliance on default configurations.

3 The authors do not specify the concrete configuration of the cloud-side DAG planner, so it cannot be determined whether the DAG construction affects the reported results. I would like the authors to describe how the cloud planner constructs the DAG and its specific configuration, and to evaluate the potential impact of these choices on the results.

Questions for authors

Q1: What evidence shows that U_i⋅D_i^sem is a valid signal-strength proxy in Eq. (6)? Q2: How sensitive are the results to the local sampling number M=5 and kernel bandwidth h=1? Q3: How is the cloud-side DAG constructed, and how do the planner's configuration and the resulting DAG structure affect the reported results?

3 comments

Response to the validity and fffectiveness of UiDisemU_i D_i^{\mathrm{sem}} in Eq. (6).

Thank you for raising this question. The product UiDisemU_iD_i^{\mathrm{sem}} is not an arbitrary heuristic. Let ViV_i denote the total unresolved task-relevant signal at node ii, and let αi\alpha_i denote the fraction of that signal attributable to private evidence. The effective private-dependent signal strength is therefore

Si=Viαi.S_i = V_i\alpha_i.

In Swim, UiU_i estimates ViV_i from the dispersion of repeated local executions, while DisemD_i^{\mathrm{sem}} estimates αi\alpha_i from the controlled representation displacement between the full-evidence and masked-evidence twin DAGs. Consequently,

S^i=UiDisem\widehat S_i = U_iD_i^{\mathrm{sem}}

serves as a direct estimator of the effective private signal strength. Under the signal-plus-Gaussian-noise model, the recoverable fraction is Si/(Si+σℓ2)S_i/(S_i+\sigma_\ell^2). Replacing the unknown SiS_i with its estimator S^i\widehat S_i yields the recovery factor used in Eq. (6).

Response to the sensitivity of the local sampling number MM and kernel bandwidth hh.

Thank you for this valuable suggestion. We conducted additional sensitivity experiments using representative values of the local sampling number MM and kernel bandwidth hh. Increasing MM from 3 to 7 yields only marginal improvements in utility and privacy metrics. Thus, M=5M=5 provides a better trade-off among estimation stability, task performance, privacy leakage, and computational cost.

Performance is also relatively stable across the evaluated bandwidths: h=0.5h=0.5 slightly degrades both utility and privacy metrics, whereas h=2.0h=2.0 reduces leakage at the cost of answer quality. These results support h=1h=1 as a balanced operating point and indicate that the method is not overly sensitive to moderate parameter variations. We will incorporate this analysis into §5.5 of the revised paper. All results are obtained on P-FinQA under the LLM1+SLM1 configuration.

Local Samples MMKernel Bandwidth hhF1 ↑SM ↑LaJ ↑AUC ↓MI ↓Local Latency (s)
31.0.686.724.778.529.16436.41
51.0.691.729.783.523.15857.86
71.0.693.731.785.521.15580.97
50.5.688.726.780.526.16159.47
52.0.685.722.776.518.15159.91

Response to the configuration and impact of the cloud-side DAG planner.

Thank you for noting that the cloud-planner configuration was underspecified. The planner receives only the edge-sanitized task descriptor and public information, and returns a schema-constrained JSON DAG specifying each node’s subtask, operation, predecessors, and input bindings. Our default implementation uses Tplan=0T_{\mathrm{plan}}=0, a 1,000-token output limit, and at most three generation-and-validation attempts. Duplicate-node and cyclic graphs are rejected, and the validated DAG remains fixed during downstream allocation and execution. We will add these details to the revision.

We further evaluate the two decoding parameters that most directly affect DAG construction: planner temperature and maximum output length. We vary one parameter at a time while fixing the planner prompt, validation procedure, executor, allocator, and privacy configuration. All results reported below are obtained on P-FinQA under the LLM1+SLM1 configuration.

ConfigurationTplanT_{\mathrm{plan}}Max tokensValid DAG (%)NodesDepthF1↑SM↑LaJ↑AUC↓MI↓
Default0.0100099.46.33.2.691.729.783.523.158
Temperature 0.30.3100099.06.43.2.689.728.781.524.159
Temperature 0.70.7100097.86.63.4.684.724.777.527.163
Maximum output 5120.050095.65.52.9.681.719.774.537.173
Maximum output 20000.0200099.56.43.2.692.730.784.524.159

Increasing temperature mainly increases DAG variability and slightly reduces the valid-DAG rate. The 500-token limit occasionally truncates or simplifies the DAG, causing a small quality reduction, whereas increasing the limit to 2,000 tokens provides no consistent benefit. Across all configurations, answer-quality differences remain within approximately one percentage point, and AUC/MI change only marginally. These results indicate that the reported P-FinQA performance is stable under moderate changes to the planner decoding configuration and support deterministic decoding with a 1,000-token limit as the default.

Reviewer_WsLUOpenReview ↗

Response to Author Rebuttal

Thank you for the rebuttal and the additional experiments. My concerns on the sensitivity to the local sampling number MM and kernel bandwidth hh, and on the configuration of the cloud-side DAG planner, are largely resolved by the new results and clarifications.

However, my concern on the validity of UiDisemU_i D_i^{\mathrm{sem}} in Eq. (6) is only partially addressed. The decomposition Si=ViαiS_i = V_i \alpha_i gives the estimator a clearer interpretation and is a reasonable theoretical motivation, but it remains a post-hoc justification: no direct empirical evidence (e.g., an ablation against alternative proxies such as UiU_i or DisemD_i^{\mathrm{sem}} alone) is provided to show that this specific formulation is effective or superior.

Overall, the rebuttal is responsive, but since my original score already reflected an inclination to accept and the above reservation is not fully eliminated, I keep my original score.

Thank you for the careful follow-up and for acknowledging that the new sensitivity and planner analyses address the corresponding concerns. We also appreciate the remaining question about whether the multiplicative proxy in Eq. (6) is empirically preferable to alternative formulations.

To address this point, we conduct a controlled proxy ablation in which the data split, reasoning DAGs, local sampling configuration, kernel bandwidth, privacy budget, cloud planner, and random seed are kept fixed. We change only the signal term in Eq. (6), comparing UiU_i, DisemD_i^{\mathrm{sem}}, their arithmetic mean, and their product.

Signal proxyF1 ↑SM ↑LaJ ↑AUC ↓MI↓
UiU_i0.3410.5950.6580.6250.198
DisemD_i^{\mathrm{sem}}0.4060.6750.6940.5990.192
(Ui+Disem)/2(U_i+D_i^{\mathrm{sem}})/20.5490.7020.7390.5610.175
UiDisemU_iD_i^{\mathrm{sem}}0.6910.7290.7830.5230.158

The product achieves the best result on all five metrics. This direct comparison complements the existing fixed-signal ablations in §5.4: those ablations show that the node-specific information carried by both UiU_i and DisemD_i^{\mathrm{sem}} is necessary, while the new proxy ablation isolates how the two signals should be combined.

The result is also consistent with the intended semantics of Eq. (6). A recoverable private-dependent signal should be large only when the node is both unresolved (UiU_i is large) and dependent on private evidence (DisemD_i^{\mathrm{sem}} is large). A single factor cannot express this conjunction, while the arithmetic mean can remain large when only one condition holds. The product enforces both conditions simultaneously.

We hope that this additional evidence addresses the remaining concern regarding the effectiveness and superiority of the formulation in Eq. (6).

Reviewer_7P2sOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work (SWIM) allocates a privacy budget across a reasoning DAG, which chooses local execution, public cloud execution or protected cloud releases. It does this by scoring actions using local uncertainty, importance, and private-evidence dependency; then it reserves budget for downstream nodes. The experiments improve answer quality and reduce empiracal leakage on privacy-augmented tasks.

Critical strengths

I think that step-level privacy allocation is a quite meaningful problem, and different reasoning steps can benefit differently from private evidence and cloud computation. Also, the fact that local candidate evaluation can avoid speculative privacy releases is pretty interesting given that it can be a sensible design choice if the observable selection process is itself protected.

Critical weaknesses

The paper didn't show the implemented routing policy to satisfy the transcript-level DP guarantee.

Also, the empirical is incomplete and inconsistent; Talbe 1 and 2 disagree on P-MultiHopRAG, and Appendix E.2 explicitly describes Table 10's leakage values as placeholders.

Also, the paper is extremely hard to follow. It uses dense, non-standard jargon and in parts reads like an experiment log rather than a conference publication. I strongly suspect that the text was largely or wholly generated by an AI model. While I generally would not consider presentation issues a critical weakness of a paper, in this case I think the presentation significantly impacts the potential value of the work.

Questions for authors

I have no questions that'd impact my assessment.

2 comments

Response to the “incomplete/inconsistent empirical evaluation” and presentation concerns.

We disagree with the reviewer’s characterization. The review elevates two localized editorial errors into the much broader claim that the empirical study is incomplete or unreliable. The evidence in the manuscript does not support that conclusion.

Tables 1–2 inconsistency. We acknowledge a transcription inconsistency between Tables 1 and 2. The two tables report different views of the same underlying evaluation, and the inconsistency is confined to a single entry rather than reflecting conflicting experiments, protocols, or datasets. We will correct the entry and the corresponding textual description. This correction does not change any method ranking, aggregate result, comparative finding, or substantive conclusion of the paper.

Table 10 wording. The leakage values reported in Table 10 are genuine experimental results, not placeholders. The sentence referring to them as placeholders is an obsolete authoring note that was inadvertently retained during revision. We will remove it and clarify the table description. This is a presentation error concerning the status of the reported values, not evidence that the experiments were unfinished or that the results were fabricated or provisional.

Empirical validity. We take responsibility for both editorial mistakes. However, they call for two precise corrections—not the wholesale dismissal of an empirical evaluation spanning three datasets, multiple LLM–SLM configurations, seven baselines, complementary answer-quality and empirical-leakage metrics, controlled ablations, sensitivity analyses, attacker-strength analysis, overhead measurements, and a real-device demonstration. Neither issue affects the experimental protocol, the underlying results, or the paper’s principal findings.

Readability and organization. We acknowledge that the manuscript is technically dense, and we will improve its presentation by adding a terminology overview, simplifying several long passages, and strengthening the roadmap connecting the theoretical and algorithmic components. Nevertheless, describing the paper as an “experiment log” is not an accurate characterization of its structure. The manuscript follows a conventional progression from problem formulation and theoretical analysis to method design, experiments, and reproducibility details. In particular, “Operational Privacy Dependency” is not unexplained terminology: it is formally defined, connected to a privacy–utility lower bound and an end-to-end composition result, and then instantiated in the proposed allocation algorithm. We would welcome specific examples of passages that remain unclear, as those would allow us to make targeted improvements.

AI-authorship speculation. The suggestion that the manuscript was “largely or wholly generated by an AI model” is unsupported. The manuscript was not largely or wholly AI-generated. We explicitly disclosed the limited uses of generative tools in the AI Use Statement, independently verified all assisted material, and take full responsibility for every claim, proof, experiment, and reported result. A stylistic impression is not evidence about authorship or scientific validity. We respectfully request that the work be assessed on its identifiable technical claims, methodological soundness, and reproducible evidence rather than speculation about provenance.

The reviewer has identified one transcription inconsistency and one obsolete drafting sentence, both of which we acknowledge and will correct. Neither affects the validity of the experiments or the conclusions drawn from them. Treating these localized and readily correctable presentation issues as grounds for a clear reject is therefore disproportionate to their actual technical significance.

I thank the authors for thoroughly engaging my reviews. Taking everything into consideration, I am keeping my score.

Your reply to me does not address the DP guarantee, and your reply to Reviewer bVKp confirms the gap: routing and gating depend on private evidence, are visible to the cloud, and are not charged to the accountant. The formal claim does not hold for the evaluated system.

Your new table also shows that most of the gain comes from privacy-free calls (0.205, against 0.214 to 0.239 for protected releases).

The table explanations are plausible, but the work requires substantial revision beyond just writing. My score rests on these points, not on how the text was produced.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.