acceptodds
Under review as a conference paper at ICLR 2027

Ask Without Telling: Local SLMs Consult Cloud LLMs Without Revealing Task Intent

Abstract

As local small language models (SLMs) increasingly collaborate with more capable cloud large language models (LLMs), a natural privacy question arises: Can a local SLM obtain cloud LLM guidance while protecting user privacy? Existing privacy-preserving SLM-LLM frameworks primarily hide sensitive values while preserving task semantics, which can still expose what the user is trying to accomplish. For example, allocating scarce medical supplies across hospitals may signal an emerging public-health emergency, while rebalancing an investment portfolio may reveal a private investment strategy, even when names and numerical values are hidden. Recent decoy-based methods further obscure task intent by hiding the real request among alternatives, but stronger protection relies on more decoys or semantic abstraction, increasing overhead or risking utility loss. More fundamentally, existing work does not systematically characterize the components of private task intent or how each should be protected. We therefore introduce task-private consultation, which characterizes task intent through two components: task context and task operation. To the best of our knowledge, this is the first systematic study of these components and their individual and joint protection in local-cloud SLM-LLM consultation. To realize this setting, we propose PriCon, an end-to-end framework that transforms the task itself through recoverable mathematical reformulation rather than hiding it among alternatives. A local closed-loop refinement mechanism further maintains privacy and recoverability throughout consultation. Experiments on 100 tasks show that PriCon reduces cloud-side task-intent inference Hit@1 to nearly 0%, versus 93-99% under sensitive-value removal and 3-30% under decoy-based protection, while preserving cloud-assisted utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
SLM-LLM Collaboration, Task-Private Consultation, Task Intent, Intent-Obscuring Rewriting

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

This paper investigates task intent leakage when a local small language model consults a cloud-based large language model. PriCon decomposes task intent into task context and task operation. It first constructs a domain-agnostic representation, then introduces alternative mathematical formulations according to the protection scope, and finally generates task narratives across different domains. The cloud model returns a solution blueprint, which the local model reconstructs and executes using the retained mappings. Experiments on 100 tasks show that Hit@1 is 0% when the context, operation, or both are jointly protected. Meanwhile, all model–setting combinations reported in Table 3 achieve a 100% subtask success rate.

Critical strengths

NA

Critical weaknesses

1.The key method is not established as a verifiable general procedure, and the specific reduction example is insufficient to substantiate the effectiveness of a cross-task framework.

The core of PriCon relies on a local model to construct an alternative mathematical formulation, establish a recovery mapping MM, and determine whether privacy is preserved and the original task remains recoverable. Section 4 explains the objectives of these steps, but the main text does not sufficiently specify how candidate formulations are constructed, how mappings are validated, or how failures are identified for general tasks. Section 4.2 requires a local reviewer to assess whether the original intent can be inferred and whether the solution can be recovered, yet it does not establish the reliability of these assessments themselves. Consequently, the critical conditions determining whether the method works are still delegated to unverified local judgments. While local trust can define a trust boundary, it does not automatically guarantee the correctness of the local model's mathematical transformations or privacy assessments.

The hierarchical DAG example in Section 4.1.2 is valid under the assumptions that different supply types are independent, integer allocations are bounded, and the objective function is additive. However, it is a dynamic programming representation tailored to a specific resource allocation problem. This example neither provides a common transformation rule applicable to all tasks nor explains how reliable mappings can be obtained for tasks lacking such a structure. Section 6 already acknowledges the difficulty of finding alternative formulations for complex operations, yet the main text still fails to clearly characterize the supported task classes and the boundaries of applicability. Reporting results across 10 task families cannot substitute for a concrete description of the transformation process and its correctness validation. Therefore, the current evidence does not establish that a locally feasible reduction example can serve as a trustworthy, general-purpose consultation method.

The technical contribution is also not adequately isolated. The Math-only and Semantic ablations merely demonstrate the changes resulting from adding an alternative narrative; they do not identify the independent value of the mathematical intermediate representation over direct, semantically rewritten formulations that remain recoverable. Since end-cloud transformation and local recovery paradigms already exist, the paper has yet to demonstrate the additional privacy benefits and gains in applicability attributable to its proposed mathematical stage.

  1. The privacy measurements, baseline coverage, and task correctness lack sufficient justification, leaving the experiments unable to substantiate the claimed privacy–utility advantages.

The privacy metrics are too coarse-grained and lack direct verification of sensitive information recovery. Table 2 relies primarily on Hit@1 and Hit@5 for the scenario, operation, and combined categories as evidence of privacy protection. The available materials do not provide sufficiently detailed label granularity or semantic scoring criteria for independent verification. Nor do they explain how partially recovered sensitive facts, objectives, and constraints are counted as leakage when the complete label is not retrieved. Existing studies provide more specific evidence of information recovery. InferDPT reports token recovery attack success rates, top-kk embedding inversion results, and n-gram leakage in cloud-generated outputs. PAPILLON calculates leakage based on annotated units of private information. EmojiPrompt compares attack-reconstructed texts with the originals using ROUGE, METEOR, and semantic similarity to examine meaning leakage that may occur even when the original wording is not recovered. The present paper reports neither entity-level leakage from the original text nor token-level recovery, and it does not establish an analysis of sensitive semantic unit recovery that covers these risks. Although intent protection requires evaluation at a higher semantic level than lexical matching, this does not eliminate the need to examine partial recovery and residual sensitive values. Failure to retrieve a coarse-grained label cannot automatically be interpreted as successful privacy protection.

The paper's own results also raise concerns about the interpretation of its metrics. Section 4.1.1 specifies that scenario protection preserves the operation, yet the operation Hit@1 scores in Table 2 drop from 98%, 95%, and 99% to 31%, 59%, and 56%, respectively. The paper must rule out the possibility that the underlying computational objective has been recovered but incorrectly scored because the original domain was not explicitly reproduced. Direct cloud consultation provides the original task to the cloud, yet its Hit@1 remains below 100%, indicating that inference failures can occur even when the information has already been disclosed. For a mechanism that deliberately constructs alternative narratives, the current evidence does not establish protection against inference by an adversary aware of the defense mechanism. Nor can zero hits be interpreted as zero leakage.

The existing comparisons do not adequately cover complete end-cloud privacy-preserving methods. The main results compare direct consultation, sensitive-value removal, and one decoy baseline. They do not report results for complete transformation–cloud processing–local recovery methods such as InferDPT, PAPILLON, and PRIV-QA, nor do they cover the decomposition-based obfuscation approach of ConfusionPrompt. Removing sensitive values once is not representative of the privacy protection and recovery capabilities of these methods. Although the native privacy objectives of different approaches would need to be harmonized before a fair comparison could be made, the current experiments do not establish such a comparison. Consequently, they are insufficient to support a broad claim of privacy–utility superiority and cannot determine whether the reported gains stem from mathematical reformulation or from the overall recovery pipeline.

The reported 100% utility lacks a reliable foundation in task correctness. The publicly available regular-expression task claims that the regular expression a*b.c+ should match aaabc, although it does not actually match that string. The two task families involving budget allocation and cross-domain composition contain 20 tasks in total that require net present value (NPV) calculations, yet neither the discount rate nor the cash-flow evaluation horizon is specified. The hospital budget allocation task does not even provide project cash flows. Lines 402–403 use direct cloud consultation as the correctness reference without establishing how errors in the reference outputs or incomplete inputs are handled. Consequently, the uniformly reported 100% utility in Table 3 cannot be verified from the public materials as evidence of functional correctness on the original tasks. It therefore cannot serve as a reliable basis for the claim that PriCon preserves privacy without sacrificing utility.

Questions for authors

1. Does the core reformulation have clearly defined applicability conditions and validation criteria? Beyond the known allocation examples that can be reduced, what establishes the correctness of the alternative representations and the mapping MM? What independent evidence substantiates the local reviewer’s judgments regarding privacy and recoverability? Does the current cross-task effectiveness fundamentally depend on task-specific transformation knowledge?

2. How does Hit@k account for partial recovery of sensitive information? How does the current scoring mechanism handle residual original entity mentions, token-level recovery, synonymous semantic recovery, and the recovery of sensitive targets and constraints? Why does the hit rate decline substantially even when operations are preserved by definition? This question directly concerns whether Table 2 measures the recovery of sensitive operations or merely the reproduction of the original narrative.

3. What evidence establishes the independent advantages of mathematical reformulation over existing approaches? Neither comprehensive end-to-cloud protection nor local recovery methods are included in the main results, and the current ablation studies do not isolate the benefits of the mathematical reformulation stage relative to direct semantic rewriting. How does the existing evidence rule out the effects of the complete recovery pipeline and additional reasoning?

4. How do deficiencies in the public tasks affect the success criteria reported in Table 3? How are incorrect expected regular expressions and missing required inputs in 20 NPV tasks handled during actual evaluation? What actual inputs and scoring criteria substantiate the reported 100% functional success rate?

4 comments
AuthorsOpenReview ↗

Response to Reviewer MAKn (Part 1/4)

We sincerely appreciate your thorough review and the many detailed and constructive comments. We have carefully considered your concerns and address them point by point below.


W1 & Q1: Automatic Reformulation and Validation

We understand your concern that a single DAG example is not sufficient to demonstrate a general task transformation method. In fact, this example is intended to illustrate how a recoverable alternative formulation can be constructed, rather than provide a universal reduction rule. The key idea behind PriCon is to automatically construct and refine recoverable surrogates, allowing local SLMs to obtain cloud guidance while protecting the original task intent.

To better clarify the key ideas and technical contributions of PriCon, let us briefly explain how the framework works.

1. Automatic reformulation and review.

As illustrated in Fig. 2, PriCon uses mathematical reformulation and semantic rewriting to transform the original task into an intent-obscuring surrogate. The local SLM automatically constructs the surrogate and its recovery correspondence M\mathcal{M}, without requiring manually designed transformations for each task.

A key challenge is ensuring that the rewritten task both protects privacy and remains recoverable. To address this, we introduce a local reviewer that checks whether the original intent is hidden and whether the necessary variables, constraints, dependencies, and recovery mapping are preserved. If either check fails, PriCon automatically revises the surrogate and repeats the review. The specific review criteria, inputs, outputs, and prompts are detailed in Appendix B.2.

2. Closed-loop verification and debugging.

Passing the review does not necessarily mean that the task can already be solved successfully. After cloud consultation, PriCon uses the private correspondence M\mathcal{M} to recover the cloud guidance and verifies the result against the original task. If execution fails, it generates a privacy-preserving diagnostic request and seeks further cloud assistance.

In this way, PriCon forms an automated, closed-loop refinement process without human intervention. Rather than solving the original task independently, the local SLM focuses on protecting task intent and recovering cloud guidance. When difficulties arise, it can seek further assistance from the cloud LLM, allowing it to progressively solve complex tasks beyond its standalone reasoning capability.

The local experience bank is optional and primarily improves efficiency, as evaluated in Section 5.3. It is disabled in our main privacy evaluation (Section 5.2), so the reported privacy results do not depend on manually prepared transformation templates or retrieved experiences.

3. Experimental evidence and why it works.

As highlighted in the Abstract, PriCon reduces cloud-side task-intent inference Hit@1 to nearly 0% while preserving cloud-assisted utility. Specifically, PriCon-Joint achieves 0% joint Hit@1 across 100 tasks and three local SLMs (Table 2). Following suggestions from other reviewers, we further evaluated mechanism-aware attacks under Kerckhoffs's principle. PriCon-Joint maintains 0% joint Hit@1 and Hit@5 even when the attacker knows the rewriting mechanism (Table 6). We also added 50 tasks from 10 new task families and evaluated DeepSeek-V4-Pro and Qwen-3.5 as attacking LLMs, obtaining 0% joint Hit@1 and at most 2% joint Hit@5 across both standard and mechanism-aware settings (Appendix D.1, Table 8).

Why does it work? PriCon replaces the original task with a coherent alternative story, shifting the attacker's inference toward the surrogate rather than the original context. Under joint protection, PriCon also changes the apparent mathematical operation. Since similar mathematical structures can arise from many application domains, the attacker faces a wide range of possible original context-operation combinations, even when the rewriting mechanism is known. Meanwhile, the private correspondence M\mathcal{M} allows the local SLM to recover and execute useful cloud guidance. This explains why PriCon can achieve strong task-intent protection across diverse task families while preserving cloud-assisted utility.

We hope these explanations and additional results help clarify PriCon's design and effectiveness. The overall workflow is illustrated in Fig. 4 (Page 7) and described in Section 4.2. To make the implementation easier to follow, we have also provided detailed procedures, including module inputs, outputs, prompts, and validation criteria, in Appendix B.

AuthorsOpenReview ↗

Response to Reviewer MAKn (Part 2/4)


W2-1: Privacy Metrics and Task-Intent Evaluation

We clarify two key points regarding our privacy evaluation. First, Hit@k is a task-intent-level metric, measuring the attacker's ability to reconstruct the original task context and operation, rather than token-level leakage or individual sensitive-value exposure. Second, the decrease in Operation Hit@1 under context protection does not indicate a change in the underlying mathematical structure; rather, it reflects the loss of contextual clues needed to infer the original task-level operation. We elaborate on these two points below.

1. Task-intent recovery metrics. Our goal is to evaluate whether the cloud attacker can reconstruct the original task intent, rather than reproduce its exact wording. The attacker generates five ranked candidate task decompositions from the complete cloud-visible transcript. Two semantic judges compare each predicted subtask with its ground truth: the context judge checks whether the original application domain or scenario is recovered, while the operation judge checks whether the primary computation or decision objective, together with its relevant target, is recovered. Semantically equivalent expressions are accepted, but generic descriptions such as allocation or optimization alone are insufficient. Joint inference requires recovering the correct context–operation pair for the same subtask within the same candidate.

Hit@1 measures the recovery rate of the first-ranked candidate, while Hit@5 takes the highest recovery rate among the five candidates. Since tasks may contain multiple subtasks, partial recovery is reflected proportionally. We emphasize that Hit@k is a task-intent-level metric, rather than an exhaustive measure of token-level or individual sensitive-value leakage. Therefore, 0% joint Hit@1 means that the attacker fails to jointly reconstruct the original context–operation pairs under the evaluated attack, not that all residual information leakage is eliminated.

In the revised manuscript, we have clarified this metric interpretation in Section 5.1 (Privacy metrics) and provided detailed attack protocols, semantic judging criteria, and metric computation in Appendix C.

2. Why does operation Hit@1 decrease under context protection? This is an important distinction. Although context protection preserves the underlying computational structure, context and operation are not entirely independent in natural-language tasks. Domain-specific entities and application scenarios provide clues about the concrete operation and its target. Once these contextual clues are replaced, the attacker may still recognize a general computational pattern but fail to recover the original task-level operation.

For example, PriCon transforms an asteroid-orbit analysis task into a reservoir-discharge analysis task:

SubtaskOriginal SubtaskRewritten SubtaskAttacker's Actual Guess
1Convert celestial coordinates into decimal degrees.Convert reservoir crest-height readings into decimal feet.Compute decimal crest heights from gauge readings.
2Calculate angular displacement and angular velocity.Compute the total change in stored reservoir volume.Compute total volume change over 20 days.
3Compute tangential velocity from angular velocity and observation distance.Compute average daily volume change and related outflow.Compute average daily volume change.
4Estimate the orbital semi-major axis using the orbital-energy equation.Estimate peak reservoir outflow capacity using a hydraulic equation.Estimate peak outflow capacity using the orifice equation.

As shown above, the attacker produces plausible reservoir-related guesses for all four subtasks, but these differ substantially from the original astronomical computations. For example, estimating reservoir volume changes or peak outflow capacity does not recover the original calculations of angular displacement, angular velocity, or orbital semi-major axis. Thus, the decrease in Operation Hit@1 does not mean that context protection changes the underlying mathematical structure; rather, it reflects the loss of contextual clues needed to infer the original task-level operation.

Thank you for pointing this out. We have added a discussion of this phenomenon in Section 5.2 (Page 8, Task-intent protection) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer MAKn (Part 3/4)


W2-2: Baseline Coverage and the Role of Mathematical Reformulation

Since task-intent protection in local–cloud LLM consultation remains relatively underexplored, we compare PriCon with representative existing protection strategies. Our results demonstrate that PriCon provides stronger task-intent protection while preserving cloud-assisted utility. Moreover, mathematical reformulation is not merely a rewriting step, but an important intermediate representation that enables broader protection capabilities and cross-domain experience reuse. We discuss these aspects below.

1. Baseline comparisons. Research on task-intent protection in local–cloud LLM consultation is still relatively limited. We therefore select representative methods from two related directions: PAPILLON for sensitive-value protection, and GS-Chaff and PrivShield for decoy-based obfuscation. These baselines are introduced in Section 5.1 (Page 8, Comparison settings). All methods are evaluated on the same 100-task benchmark, using the same local SLMs, cloud LLM, attacker, and task-intent metrics, under the same end-to-end consultation and local recovery setting.

For a concise comparison, the following table reports Hit@1 (%) with GPT-4o-mini, using the results from Table 2.

Inference TargetPAPILLONGS-ChaffPrivShield-5PrivShield-20PriCon-ContextPriCon-Joint
Context941310600
Operation99273018310
Joint94127300

PAPILLON, the sensitive-value-protection baseline, yields high Hit@1 scores of 94–99%, indicating that hiding sensitive values alone does not prevent the cloud from inferring task intent. In contrast, PriCon-Context reduces context Hit@1 to 0%, while PriCon-Joint further protects the operation and achieves 0% across all three inference targets.

Both decoy-based methods, GS-Chaff and PrivShield, reduce task-intent inference but still leave residual leakage. For example, even with 20 decoys, PrivShield retains 3% joint Hit@1 and 18% operation Hit@1, whereas PriCon-Joint reduces both to 0%. The other two local SLMs exhibit the same overall trend, with minor variations in the exact scores.

Meanwhile, PriCon maintains 100% cloud-assisted subtask success across all three local SLMs. As discussed in Section 5.2 (Page 9, Utility and privacy-overhead trade-off), PriCon-Joint achieves comparable token usage to PrivShield-5 while providing stronger task-intent protection. PrivShield-20 incurs approximately 4.9× the token usage of PriCon-Joint but still leaves residual joint inference.

2. Why is mathematical reformulation important? Semantic rewriting can directly transform an original task into a coherent story from another application domain. When only context protection is required, semantic rewriting can serve this purpose. Our motivation for retaining mathematical reformulation is not that semantic rewriting alone cannot protect task context, but that the mathematical representation provides an important intermediate step for broader protection capabilities and subsequent experience reuse.

As illustrated in Fig. 2, when operation protection is required, PriCon builds on this mathematical representation to construct an alternative formulation that changes the apparent computation while preserving the necessary mathematical structure for local recovery through the private correspondence M\mathcal{M}. Furthermore, as illustrated in Fig. 4, the same representation allows the local experience bank to match tasks from different domains based on their underlying mathematical structures, rather than superficial similarities between their semantic descriptions. This enables cross-domain experience retrieval and reuse.

Table 3 compares mathematical reformulation alone (Math-only) with the complete rewriting process (Semantic). For GPT-4o-mini under joint protection, Math-only reduces context and operation Hit@1 to 3% and 1%, respectively, while Semantic further reduces both to 0%, demonstrating the additional benefit of semantic rewriting. The experience bank is disabled in Section 5.2 and therefore does not affect these privacy results; its efficiency benefits through experience reuse are evaluated separately in Section 5.3.

We thank you again for this insightful observation. It has prompted us to clarify why mathematical reformulation is retained as an important intermediate step in Section 4.1.1 (Page 5).

AuthorsOpenReview ↗

Response to Reviewer MAKn (Part 4/4)


W2-3: Utility Evaluation and Benchmark Refinement

Thank you for carefully examining our utility evaluation and benchmark. We clarify the utility criteria and address the benchmark concerns below.

1. Utility evaluation and direct consultation reference. PriCon does not require the cloud LLM to return an executable file. A subtask is considered successful through either of the following two paths:

  • Direct solution: The cloud LLM directly provides the final answer or executable code, which can be checked or executed locally to obtain the expected result.
  • Local recovery: The cloud LLM provides a solution procedure or blueprint. The local SLM maps it back to the original task through the private correspondence M\mathcal{M}, incorporates locally retained values, and completes the execution.

In both cases, success is determined by verifying the locally obtained result against the corresponding task requirements and reference answers.

We use direct cloud consultation as the utility reference because the selected tasks cannot be reliably solved by local SLMs alone. It provides a natural baseline for evaluating how well PriCon preserves the utility of cloud assistance while protecting task intent.

2. Benchmark Design and Refinement. Since task-intent protection in local–cloud LLM collaboration is still an emerging research area, publicly available benchmarks remain limited. We therefore constructed this benchmark using tasks that can be solved with cloud LLM assistance but cannot be reliably completed by the three evaluated local SLMs alone, making cloud consultation necessary. As an initial effort, some task descriptions and annotations may still require refinement, and we appreciate the reviewer for helping us identify these issues.

For tasks that do not require instance-specific numerical values, utility can still be directly evaluated by checking the returned answer or executing the provided code against the task requirements, as described in the first success criterion above.

We will further review and refine the benchmark by providing complete task specifications, fixed reference answers, and corresponding verification scripts to improve its reliability and reproducibility. We hope this benchmark can serve as a useful resource for other researchers and facilitate future studies on task-intent protection.

Reviewer_7QNTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates task-intent privacy in local small language model and cloud large language model collaboration. The authors argue that existing privacy-preserving approaches primarily focus on hiding sensitive values, while the application context and requested operation of a task may still reveal private user intent. To address this issue, the paper introduces the concept of task-private consultation, characterizing task intent through two components: task context and task operation, and considering context protection, operation protection, and joint protection. Based on this formulation, the authors propose PriCon, a framework that transforms the original task into an intent-obscuring surrogate through mathematical reformulation, alternative mathematical formulation, and semantic rewriting. The framework further incorporates local review, closed-loop refinement, and a local experience bank to maintain privacy and recoverability while reducing consultation overhead through experience reuse. Experiments are conducted on 100 tasks across 10 task families, using three local SLMs paired with a cloud LLM. The results show that PriCon achieves nearly 0% Hit@1 for task-intent inference under joint protection, outperforming sensitive-value removal and decoy-based protection.

Critical strengths

Existing privacy-preserving approaches for local-cloud language-model collaboration primarily focus on protecting sensitive entities, attributes, or values. This paper highlights that task context and task operation may still reveal private user intent even when such information is hidden. By characterizing task intent through these two components and considering their coupled leakage, the paper provides a useful perspective for studying task-intent privacy.

Unlike decoy-based methods that hide the original request among alternative queries, PriCon transforms the original task into a different mathematical formulation and application context while preserving a local correspondence for solution recovery. This demonstrates the possibility of obscuring task intent while retaining the solution-relevant structure. Together with transcript-level intent evaluation and local closed-loop refinement, the framework offers a promising direction for task-intent protection.

Critical weaknesses
  1. PriCon-Joint achieves nearly 0% Hit@1 in Table 2. However, the current evaluation does not clearly specify whether the attacker is aware of PriCon's mechanisms, rewriting strategies, or the candidate set of original tasks. The low inference accuracy may primarily reflect the distribution shift introduced by semantic rewriting rather than the inherent difficulty of recovering the original task intent. Following Kerckhoffs's principle, the authors should further evaluate privacy protection under an attacker who knows the protection mechanism to strengthen the credibility of the results.

  2. PriCon also introduces substantial additional computation and communication overhead compared with direct cloud consultation. Its LLM token usage is several times higher, and the number of consultation rounds increases considerably. Without sufficient evidence of additional utility gains, this overhead further raises concerns about the practical value of the framework.

  3. The evaluation includes only 100 tasks, which is relatively limited for demonstrating the generalizability of the proposed framework. Furthermore, although the paper claims to provide a new systematic characterization of task-intent privacy, its core technical approach still largely relies on task rewriting and mathematical reformulation. Thus, the novelty of the contribution may be somewhat overstated.

  4. Although the paper discusses its applicability boundaries, it does not systematically investigate which types of tasks can be reliably reformulated and which cannot. A more concrete evaluation of the range of applicable tasks is also important for establishing the practical necessity of the proposed approach. Therefore, the applicability and limitations of PriCon remain insufficiently explored.

  5. Some figure titles and labels are too small, affecting readability. The authors are encouraged to check whether the formatting fully complies with the conference template requirements.

Questions for authors

In the task-intent inference experiments, is the attacker aware of PriCon's mechanisms, rewriting strategies, and the candidate set of original tasks? Would PriCon still achieve nearly 0% Hit@1 under a method-aware attacker, consistent with Kerckhoffs's principle?

How representative are the 100 benchmark tasks of the broader task-private consultation setting? Furthermore, could the authors clarify what distinguishes the proposed systematic characterization of task intent from existing task rewriting and semantic abstraction approaches beyond the combination of these techniques?

Although Section 6 discusses applicability boundaries, could the authors provide a more systematic analysis of which task families can be reliably reformulated and which cannot? In particular, what are the reformulation failure rates across different task families?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7QNT (Part 1/2)

We appreciate your thoughtful comments and suggestions. Our responses are provided below.


W1 & Q1: Protection under Mechanism-aware Inference

This is an important concern. We agree that low Hit@1 in the original setting alone does not establish protection against an informed attacker. To address this, we have added a stronger evaluation following Kerckhoffs's principle.

1. Mechanism-aware attack. In the original setting, the honest-but-curious cloud infers task intent from the complete cloud-visible transcript without being explicitly informed of PriCon's rewriting mechanism. The attacker is not given a predefined set of possible original tasks. Instead, it generates five ranked candidate reconstructions, each with the specified number of subtasks.

Following your suggestion, we further evaluate a mechanism-aware inversion attack, where the attacker receives both the complete transcript and PriCon's rewriting, review, consultation, execution, and debugging procedures. It attempts to reconstruct the original task, while sensitive values and the task-specific correspondence M\mathcal{M} remain private.

As shown in Table 6 (Page 10), PriCon-Joint maintains 0% joint Hit@1 and Hit@5 across all three local SLMs. In comparison, PAPILLON, which protects sensitive values, yields 88–92.2% joint Hit@1, while the decoy-based GS-Chaff still yields 15.8–24.8%.

Why does it work? Semantic rewriting replaces the original context with a coherent alternative story, while PriCon-Joint also changes the apparent mathematical operation. Since similar mathematical structures can arise from different application domains, the attacker still faces many possible original context-operation combinations, even when the rewriting mechanism is known. Meanwhile, the private correspondence M\mathcal{M} allows the local SLM to recover and execute useful cloud guidance.

2. Additional evaluation with different attacking LLMs. To further examine whether this protection extends to different tasks and attacking models, we evaluated 50 new tasks from 10 additional task families using DeepSeek-V4-Pro and Qwen-3.5 as attacking LLMs. Under both transcript-only and mechanism-aware settings, PriCon-Joint achieves 0% joint Hit@1 and at most 2% joint Hit@5, further demonstrating its effectiveness across diverse tasks and its robustness to mechanism-aware inference. Refer to Appendix D.1 (Table 8) for full results.

To make these points clearer, in the revised manuscript, we have updated the attacker knowledge description in Section 3.3 (Page 5), expanded the rewriting rationale in Sections 4.1.1 (Page 5) and 4.1.2 (Page 7), and added the mechanism-aware evaluation in Section 5.2 (Pages 9–10, Table 6). These changes are highlighted in blue in the revised manuscript. We have also provided detailed attack procedures in Appendix C.2.


W2: Practical Utility and Privacy–Overhead Trade-off

This is a fair concern. Direct cloud consultation has lower overhead in our experiments, but it exposes the original task and its intent to the cloud. Moreover, when many task instances share the same underlying algorithm, repeatedly sending their data to the cloud can lead to substantial token and communication costs. PriCon instead obtains reusable cloud guidance and performs task-specific execution locally.

We recognize that protecting task intent introduces additional overhead, but we believe this trade-off is worthwhile given the privacy gains. As shown in Tables 2 and 4 (Page 9), PriCon-Joint requires about 4.4× more LLM tokens than direct consultation, while reducing joint Hit@1 from 93–98% to 0% and maintaining the same 100% subtask success rate.

Compared with PAPILLON, which protects sensitive values, PriCon-Joint reduces joint Hit@1 from 93–96% to 0%, at the cost of 6.8× more tokens. Compared with decoy-based protection, PriCon-Joint uses comparable tokens to PrivShield-5 while reducing joint Hit@1 from 6–7% to 0%. Even with 20 decoys, PrivShield requires 4.9× more tokens than PriCon-Joint, yet retains 3–4% joint Hit@1. These results highlight PriCon's favorable privacy-overhead trade-off.

To better clarify these comparisons, we have expanded the discussion in the revised manuscript. Please see Section 5.2 (Page 9, Utility and privacy-overhead trade-off) for details.

AuthorsOpenReview ↗

Response to Reviewer 7QNT (Part 2/2)

W3 & Q2: Generalizability and Methodological Novelty

1. Generalizability across task families. Since task types in real-world applications are highly diverse, it is difficult to cover all possible tasks. We therefore selected 10 distinct task families, with 10 instances each, covering structured reasoning, decision-making, optimization, coding, and cross-domain tasks. Importantly, the selected tasks cannot be reliably solved by local SLMs alone, motivating the need for cloud consultation.

To further address your concern, we have added 50 new tasks from 10 additional task families, covering scientific, language, procedural, and security-related reasoning. Consistent with our original findings, PriCon-Joint maintains 0% joint Hit@1 and at most 2% joint Hit@5 across different attacking LLMs and settings. As discussed in our response to W1 & Q1, these additional results further support PriCon's effectiveness across diverse task families. Full results are provided in Appendix D.1 (Tables 7 and 8).

2. Methodological novelty. We believe these consistent results reflect PriCon's distinct privacy objective and technical design, which go beyond conventional task rewriting.

Existing sensitive-value protection methods mainly hide private entities or numerical values while preserving task semantics, which can still reveal user intent. Decoy-based methods instead hide the original request among alternatives, introducing additional overhead while leaving potential intent leakage.

PriCon takes a different perspective. We first systematically characterize task intent through task context and task operation, analyze their coupled leakage, and define corresponding protection scopes. This characterization directly guides a simple yet effective approach: PriCon constructs a recoverable alternative formulation that selectively transforms the apparent application context, mathematical operation, or both, while keeping the recovery correspondence M\mathcal{M} private. This enables more targeted and efficient task-intent protection while preserving cloud-assisted utility.

Importantly, PriCon implements this transformation through an automated, closed-loop workflow without human intervention. Failed privacy or recoverability reviews trigger further rewriting, while execution failures lead to privacy-preserving debugging and additional cloud assistance. The local SLM therefore does not need to solve the original task independently. Instead, it focuses on protecting task intent and recovering cloud guidance, allowing it to benefit from reasoning capabilities beyond its own.

Taken together, we believe PriCon makes contributions at both the problem and method levels. It provides a systematic understanding of task-intent privacy and a practical, recoverable mechanism for protecting it, enabling resource-constrained local SLMs to leverage powerful cloud LLMs without directly revealing their original tasks.

The protection framework is detailed in Section 3.2 (Page 4), while the technical mechanisms are described in Sections 4.1–4.2 (Pages 5–7) and Appendix B.


W4 & Q3: Applicability and Reformulation Reliability

This is an important point. As discussed in our response to W3 & Q2, we have expanded our evaluation to 20 distinct task families, covering a broad range of reasoning and computational tasks (Appendix D.1). These tasks share an important property: their underlying operations can be separated from the original application context and transformed into recoverable alternative formulations.

For such tasks, PriCon uses an automated, closed-loop refinement process without human intervention. Failed privacy or recoverability reviews trigger further rewriting, while execution failures lead to privacy-preserving debugging and additional cloud assistance. This allows PriCon to progressively refine the surrogate and recover useful cloud guidance without relying solely on the local SLM's reasoning capability.

However, we recognize that PriCon is less suitable for tasks strongly dependent on domain-specific knowledge, such as medication-related queries. Their answers inherently depend on the original context, making it difficult to construct a valid alternative formulation. A promising direction is to maintain domain knowledge locally and combine local knowledge retrieval with PriCon's rewriting mechanism. However, this requires a more comprehensive investigation. We appreciate your valuable comment, which encourages us to explore more diverse mechanisms to further broaden PriCon's applicability in future work.

We have discussed these applicability boundaries in Section 6 (Page 10, Applicability boundary).


W5: Formatting

Thank you for pointing this out. We have enlarged the figure titles and labels to improve readability and carefully checked the manuscript formatting to ensure compliance with the conference template requirements.

Thank you to the authors for their thoughtful response. My remaining concerns primarily relate to the additional overhead during execution, such as the 4.4× or 6.8× increase in cost. For scenarios involving frequent use, the framework not only requires the deployment of a local SLM but also introduces additional execution overhead, which somewhat limits its applicability. Unless there is an urgent need for privacy protection, these additional costs may be difficult to justify.

Although I still have some reservations regarding this issue, considering the additional explanations and clarifications provided by the authors on other aspects, I am willing to raise my final score to 3.

AuthorsOpenReview ↗

Response to Reviewer 7QNT

Thank you for reconsidering your assessment. We agree that the added overhead matters when deciding whether to use PriCon. When multiple privacy-sensitive task instances share the same underlying algorithm, PriCon can reuse guidance from the initial cloud consultation and run subsequent instances locally, spreading the consultation cost across runs.

For new tasks, the local SLM’s limited reasoning ability can lead to additional debugging rounds before a solution runs successfully. These rounds account for much of the added token cost. Figure 5 shows that reusing rewriting and debugging experience from similar tasks can reduce token use in PriCon-Joint. Inspired by your comments, we will refine the debugging process to reduce the number of cloud consultation rounds needed for new tasks. Thank you again for your time and valuable comments.

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

This paper studies task-private consultation between a trusted local SLM and a more capable cloud LLM. It separates task intent into context and operation and considers context-only, operation, and joint protection. PriCon constructs a recoverable surrogate through mathematical reformulation and semantic rewriting while retaining the correspondence and sensitive values locally. The cloud supplies a solution blueprint, which the client maps back and executes. Local review, surrogate-level debugging, and an experience bank support repeated consultation. A 100-task benchmark evaluates intent inference from the complete cloud-visible transcript, cloud-assisted utility, and consultation overhead.

Critical strengths
  1. The decomposition of task intent is useful. Application context and requested operation are distinct privacy targets, and contextual cues can reveal an operation even when its explicit description is changed. The three protection scopes and the coupling analysis make this distinction operational rather than merely terminological (Section 3.2 and Table 5). This broadens the discussion beyond removing names and numerical values while retaining the user's underlying purpose.
  2. The method goes beyond entity substitution. The medical-supply allocation example is first abstracted mathematically and then reformulated as shortest-path search on layered DAGs, with a local correspondence between path choices and allocations. This gives a concrete reason why the cloud's guidance can remain useful after the apparent task has changed (Sections 4.1.1-4.1.2). The recoverable alternative formulation is the most technically interesting part of the proposal.
  3. The same example is carried through the major design stages. Moving from the original allocation task to a neutral mathematical representation, an alternative formulation, a routing story, and local recovery makes it easier to see what each operation contributes. The treatment of local review and debugging also helps explain how the interface is intended to function over a complete consultation rather than a single sanitized request (Figures 2-4).
Critical weaknesses
  1. Low intent-inference Hit@1 does not establish protection against an informed adversary. It is unclear whether the attacker knows PriCon's transformations and attempts to reverse them; simply following the surrogate story could explain low recovery. Kerckhoffs's principle [1] calls for public-design knowledge while keeping the request-local mapping M and private values secret. An honest-but-curious cloud can analyze transcripts offline, including alternative original-task hypotheses and auxiliary knowledge. Evaluate method-aware inversion and, where relevant to deployment, cross-session information, or explicitly exclude it. Current evidence supports protection against the tested inference procedure; robustness to this stronger knowledge model remains open.
  2. Competitive comparisons are limited. Section 5.1 includes direct consultation, sensitive-value removal, and one PrivShield decoy-based baseline; five versus twenty decoys are operating points of that baseline. GS-Chaff [2], already discussed in Section 2, combines semantic abstraction with chaff generation but is not evaluated. An additional comparison should isolate recoverable reformulation from stronger rewriting/refinement, using the same cloud model, comparable local resources, and matched token/call budgets while distinguishing protection goals.
  3. Recoverable reformulation limits applicability to tasks whose domain context can be separated from computation and mapped reliably to an alternative problem. Medication knowledge queries do not obviously qualify, as Section 6 acknowledges. Quantify eligible workload coverage, automatic reformulation and recovery failures, manual intervention, and fallback behavior. Constructing and checking surrogates may itself demand substantial local reasoning, which matters when cloud consultation is motivated by the local model's inability to solve the original task.
Questions for authors
  1. Is the intent attacker given PriCon's public procedure and examples of its transformation families? Could you evaluate an explicit reverse-reformulation attack with several prompts or independent models, without revealing the secret per-request mapping, and state what auxiliary or cross-session information is allowed?
  2. Could you add GS-Chaff, or explain a concrete incompatibility with its evaluation, and compare it with PrivShield and PriCon under matched call/token budgets? Please clarify how the original baseline mechanisms are adapted to the same task families and inference targets.
  3. What fraction of candidate tasks can be automatically converted into a valid recoverable surrogate, and how is correctness checked? Please report reformulation and recovery failures, human intervention, and the fallback for semantic-knowledge tasks. Distinguish new task structures from familiar templates or retrieved experience.
4 comments
AuthorsOpenReview ↗

Response to Reviewer jwkB (Part 1/2)

We are grateful for the careful review and constructive comments. We address each point below.


W1 & Q1: Mechanism-aware task inversion

Thank you for this valuable suggestion. We agree that low Hit@1 under the original attack setting alone is insufficient to establish protection against an informed attacker. We have therefore added a stronger evaluation following Kerckhoffs's principle.

1. Mechanism-aware attack. In the original setting, the honest-but-curious cloud infers task intent from the complete cloud-visible transcript, including surrogate queries, cloud responses, and debugging exchanges, without being explicitly informed of PriCon's rewriting mechanism.

Following your suggestion, we further evaluate a mechanism-aware inversion attack. The attacker receives both the complete transcript and descriptions of PriCon's rewriting, review, consultation, execution, and debugging procedures. It attempts to reconstruct the original task, while sensitive values and the task-specific correspondence M\mathcal{M} remain private.

The new results in Table 6 (Page 10) show that PriCon-Joint maintains 0% joint Hit@1 and Hit@5 across all three local SLMs, compared with 15.8–24.8% Hit@1 and 28–33.8% Hit@5 for GS-Chaff.

Why does it work? In the original setting, semantic rewriting replaces the original task with a coherent alternative story, shifting the cloud LLM's inference toward the surrogate rather than the original context. Under joint protection, PriCon also changes the apparent mathematical operation. Since similar mathematical structures can arise from many application domains, the attacker still faces a wide range of possible context-operation combinations, even when the rewriting mechanism is known. Meanwhile, the private correspondence M\mathcal{M} allows the local SLM to recover and execute useful cloud guidance.

Manuscript revisions: We clarified the attacker knowledge in Section 3.3 (Page 5), explained the rewriting rationale in Section 4.1.1 (Page 5) and Section 4.1.2 (Page 7), and added the mechanism-aware evaluation in Section 5.2 (Pages 9–10, Table 6). These changes are highlighted in blue in the revised manuscript. Detailed PriCon procedures and attack protocols are provided in Appendices B and C.

2. Additional evaluation with different attacking LLMs. We further evaluated 50 new tasks from 10 additional task families, using GPT-4o-mini as the local SLM and DeepSeek-V4-Pro and Qwen-3.5 as two attacking LLMs. Both transcript-only and mechanism-aware settings were considered. The table below summarizes joint intent inference (context-operation pair) using Hit@1 / Hit@5 (%). Full results are provided in Appendix D.1 (Table 8).

Attacking LLMMechanism-awareGS-ChaffPriCon-Joint
DeepSeek-V4-ProNo17.5 / 240 / 0
DeepSeek-V4-ProYes11.5 / 23.50 / 0
Qwen-3.5No4 / 140 / 2
Qwen-3.5Yes4 / 13.50 / 0

Across both attacking LLMs and settings, PriCon-Joint maintains 0% joint Hit@1 and at most 2% joint Hit@5. These results further demonstrate its effectiveness across diverse task families and its robustness to mechanism-aware inference.

AuthorsOpenReview ↗

Response to Reviewer jwkB (Part 2/2)

W2 & Q2: Stronger Baselines and Fair Comparisons

Thank you for this helpful suggestion. We have added GS-Chaff (Zhou et al., 2026), which combines semantic abstraction with decoy queries, and clarified that our sensitive-value protection baseline follows PAPILLON (Siyan et al., 2025).

All methods are evaluated on the same 100-task benchmark, using the same three local SLMs, cloud LLM (DeepSeek-V4-Pro), inference targets, and metrics. Both decoy-based methods, GS-Chaff and PrivShield, use 5 decoys for comparison, while PrivShield is additionally evaluated with 20 decoys to examine the privacy-overhead trade-off.

The table below summarizes joint Hit@1 (%) from Table 2 (Page 9).

Local SLMGS-ChaffPrivShield-5PrivShield-20PriCon-Joint
GPT-4o-mini12730
DeepSeek-V4-Flash14640
Qwen3.5-Flash11740

GS-Chaff reduces joint Hit@1 to 11–14%, while PriCon-Joint achieves 0% across all three local SLMs. Comparing Tables 2 and 4 (Page 9), PriCon-Joint uses comparable tokens to PrivShield-5, while PrivShield-20 requires 4.9× more tokens yet retains 3–4% joint Hit@1. PriCon also preserves 100% cloud-assisted subtask success (Appendix D.2.1).

Unlike decoy-based methods that hide the original request among alternatives, PriCon transforms it into a recoverable surrogate, achieving stronger task-intent protection with a favorable privacy-overhead trade-off while preserving cloud-assisted utility.

For more details, please refer to Section 5.1 (Page 8, Comparison settings) and Section 5.2 (Page 9, Tables 2 and 4, Utility and privacy-overhead trade-off).

W3 & Q3: Applicability and automatic reformulation

1. Automatic reformulation and recovery. PriCon automatically performs task rewriting, privacy and recoverability review, cloud consultation, and local execution without human intervention. The local SLM does not need to solve the original task independently. Instead, it focuses on protecting task intent and constructing a recoverable surrogate, while relying on the cloud LLM for the solution procedure.

To maintain privacy and recoverability, the local reviewer checks the surrogate before cloud consultation. Failed reviews trigger further rewriting, while execution failures trigger privacy-preserving debugging and additional cloud assistance. This closed-loop process allows PriCon to progressively refine the solution without relying solely on the local SLM's reasoning capability.

We have further emphasized this automated workflow in Section 4.2 (Page 7).

2. Applicability to semantic-knowledge tasks. We agree that tasks strongly dependent on domain-specific knowledge, such as medication-related queries, are more challenging. Their answers depend on the original context, so rewriting them into another domain may produce incorrect guidance.

A promising direction is to maintain domain knowledge locally and combine local knowledge retrieval with PriCon's rewriting mechanism. However, this requires a more comprehensive investigation. We appreciate this valuable suggestion and plan to explore broader task families and applicability in future work.

For convenient reference, Appendix B provides detailed module inputs, outputs, prompts, and review procedures.

The new mechanism-aware attacks and GS-Chaff comparison help, and I still find the context–operation distinction useful. What’s missing for me is how often the rewriting actually works across a broader set of candidate tasks, with a fixed limit on retries and debugging. Complete task specifications and reference answers would also make the utility results easier to check. The cost comparison could be clearer too. I’m keeping weak acceptance, but I’d keep the claims focused on the kinds of tasks tested here. Zero joint hits shouldn’t be presented as zero leakage.

AuthorsOpenReview ↗

Second Response to Reviewer jwkB

Thank you for recognizing the additional experiments. The context-operation distinction is one of the main contributions of our work, and we’re glad you found it useful.

Your follow-up questions help us clarify what the current evaluation shows and where it can be strengthened.

(1) Debugging overhead and rewriting success. The table shows the percentage of the 100 evaluated tasks that used each range of additional cloud consultation rounds for debugging. PriCon-Context protects the context, while PriCon-Joint protects both the context and the operation.

Local SLMMethod1–34–56–78–10
DS-FlashPriCon-Context0%21%40%39%
DS-FlashPriCon-Joint10%36%16%38%
GPT-4o-miniPriCon-Context1%63%25%11%
GPT-4o-miniPriCon-Joint0%24%37%39%
Qwen3.5-FlashPriCon-Context10%50%20%20%
Qwen3.5-FlashPriCon-Joint2%32%21%45%

PriCon-Joint tends to require more debugging rounds. Rewriting both the context and the operation places greater demands on the local SLM’s reasoning and can lead to more debugging on new tasks. These extra rounds account for much of the added token cost.

With repeated use, the initial cloud consultation cost can be amortized across runs. When multiple privacy-sensitive task instances share the same underlying algorithm, PriCon can reuse guidance from the initial cloud consultation and run subsequent instances locally, spreading the consultation cost across runs. Figure 5 also shows that reusing rewriting and debugging experience from similar tasks can reduce token use in PriCon.

Following your feedback and a similar suggestion from another reviewer, we’ll refine the debugging process to reduce the number of cloud consultations needed for new tasks. We’ll then evaluate a broader set of candidate tasks and report rewriting success rates under fixed limits on retries and debugging rounds.

(2) Task-intent benchmark design. Public benchmarks for task-intent protection in local–cloud LLM collaboration are still limited. We built our benchmark around tasks that can be solved with cloud LLM assistance and that none of the three evaluated local SLMs could reliably complete on its own. We’ll add reference answers so readers can verify the utility results and reuse the benchmark in future studies. The task specifications are already available in our open-source benchmark repository.

(3) Task-intent recovery metric. Hit@k measures task-intent inference at the semantic level. We report it for context and operation separately, as well as jointly. A joint Hit@1 of 0% means that, under the evaluated attack, no top-ranked prediction correctly inferred both the original task context and the intended operation. In the revised manuscript, we’ll explain how we score the attacker’s predictions, including what counts as a hit for context, operation, and joint recovery. We’ll also state our protection claims more precisely.

Thanks again for your helpful suggestions and positive assessment of our work.

Comments on the paper

1 comment
Area_Chair_CkY8OpenReview ↗

Reminder: please read the author responses

Dear reviewers,

The authors have posted responses to your reviews. Please read them and the revised submission, and let the authors know whether your concerns are addressed. If your assessment has changed, please update your review.

Thank you for your time.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.