acceptodds
Under review as a conference paper at ICLR 2027

Learning to Refer: Client-Resolved Generation for Privacy-Aware Language Models

Abstract

Cloud-based large language models (LLMs) require users to disclose plaintext data to service providers, creating privacy risks in sensitive domains. Existing privacy-preserving approaches often trade utility for protection, incur substantial computational or communication overhead, remain vulnerable to reconstruction from intermediate representations, or protect only a subset of the training and inference pipeline. We introduce Client-Resolved Generation (CRG), a genera- tion interface that separates server-side generation from the lexical realization of input-derived content. The client transmits only pooled and noise-perturbed rep- resentations, while input-derived output content is represented using request-local positional references and resolved to its original strings only on the client. This interface protects private input and input-derived output content during both train- ing and inference while allowing the service provider to keep its proprietary model parameters hidden from the client. At the same time, exact lexical reuse remains possible without directly exposing the reused content on the provider-visible gen- eration path. We evaluate CRG on medical and document-grounded QA, sensi- tive identifier transfer, and tool calling, together with reconstruction and raw-logit leakage analyses. On SealTools, CRG improves complete-call exact match from 57.3% to 79.9% over the input-privacy framework PPFT, with larger gains as more required output content can be resolved through references. Together, these results show that CRG provides a practical interface for privacy-sensitive cloud LLMs by reducing plaintext exposure across both input and output pathways while preserv- ing task utility and server-side model confidentiality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language models, privacy-preserving inference, output privacy, client-server inference, reference-based generation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_SVbuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper observes that text-free LLM inference, which sends perturbed embeddings instead of plaintext prompts, still leaks private content on the output path when responses must reproduce input strings such as medical terms, identifiers, or tool-call arguments. The proposed CRG keeps PPFT-style pooled and noise-perturbed inputs, but trains the server decoder to emit request-local reference tokens for input-derived content, which only the client resolves to plaintext. Across 13 benchmarks and four reconstruction attacks, CRG improves utility over PPFT, most clearly on tasks requiring exact reuse (SealTools call EM from 57.3% to 79.9%), while reducing output-side leakage.

Critical strengths
  1. The problem is real and well motivated: output-side leakage of input-derived content is a genuine gap in prior input-only protection.

  2. The reference mechanism is a clean technical contribution, since the decoder must learn to point to source positions without ever seeing the plaintext source tokens.

  3. The privacy evaluation is thorough, covering input, joint, output-only, and raw-logit attacks plus a prompt-injection stress test in which PPFT leaks 95% of ages while CRG stays below 1%.

Critical weaknesses
  1. The contribution is narrower than the framing suggests. The input side (client encoder, k-pooling, Laplace noise, continuous-input decoder) is inherited directly from PPFT, so the novelty lies only in the output-side reference mechanism.

  2. The baselines miss the closest prior work on output protection. HaS and PrivacyRestore, whose placeholder-and-restore design is conceptually closest to CRG, are not compared even though the privacy experiments use PrivacyRestore's data splits, so it is unclear how much of the gain comes from the reference mechanism itself.

  3. The experimental configuration is narrow. All results use a single 1B decoder, and the main utility and privacy results use a single noise level (ϵ=75), with privacy attacks only on DDx-Plus; there is no privacy-utility trade-off curve, no larger-model study, and no formal DP guarantee.

Questions for authors

Null

2 comments

Thank you for the response. I will update my score in "Soundness" from 2 to 3.

Thank you for your thoughtful feedback and for reconsidering your assessment. Your comments have helped us improve the paper.

Reviewer_1P69OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper considers the problem of using cloud-based LLMs to answer queries where both the input and output text contain sensitive data. The authors propose a framework called client-resolved generation (CRG). The high level idea of CRG is to replace tokens in the input which are sensitive with reference tokens "[word i]" with the mapping from input tokens to reference tokens being held in a local table. This replacement is done client-side, and combined with Laplace perturbations after encoding and pooling the replaced tokens. To train the model to handle the reference tokens appropriately, the encoder, projection, and decoder are trained on input-output pairs where the input-output have references replaced, and then the projection and LoRA layers are fine-tuned on domain-specific data, again with the reference replacement applied to the input-output. This gives a model whose output can use reference tokens, which the client can locally decode using their local table. Experiments show that compared to the prior works of AlienLM and PPFT, CRG achieves higher average utility across a variety of medical utility benchmarks and structured/agentic utility benchmarks. Reconstruction attacks also demonstrate that CRG achieves better empirical privacy than the prior works.

Critical strengths
  • The idea of replacing sensitive tokens with reference tokens is a natural idea, yet the work needed to design a model training system such that this idea does not result in a huge loss of utility is significant. Upon reading the idea the first time, I felt this would be catastrophic for model utility, yet the experiments demonstrate that the utility is in fact better than past work. A natural idea that is simple to describe and works effectively is a great research contribution, and at the same time the pipeline of reference-aware alignment and then domain adaption the idea requires means that the paper also constitutes a meaningful technical contribution.
Critical weaknesses
  • CRG requires doing additional domain-specific/client-side work of deciding what input tokens need to be replaced by reference tokens and which ones are safe to send to the server verbatim. I believe PPFT does not need this additional work, and it's not clear how this work compares to the reversable vocabulary transform of AlienLM, or how automatable this step is for a new domain. Even if this is a non-trivial overhead to past work and non-automatable work, showing that this work leads to better privacy and utility is a meaningful contribution. However, it is hard to understand exactly what the impact/improvement of the paper is due to this clarity issue.
  • There are some natural questions the empirical comparisons fail to answer. (1) If one runs CRG without perturbation, Tables 1 and 2 suggest large utility improvement. However, there is no privacy comparison to CRG without perturbation. It's plausible that CRG without perturbation is also a highly private algorithm and the perturbation is a somewhat vestigial design choice from PPFT, but the experiments don't give clarity on this. (2) The 'upper bound' baseline is CRG with no noise in Tables 1 and 2, which still includes privacy protections. There's no non-private baseline. So it's not clear what the effect of the main tool of the paper, replacement with reference tokens, is on the utility. Since the impact of replacing with reference tokens isn't isolated by the experiments, it's harder to distinguish the impact of the authors' new technique with techniques from past works, and thus harder to assess the overall impact of the paper.
Questions for authors
  • Does PPFT do something to protect privacy that CRG does not? From the description in Section 2.1, it seems like it just does the same Laplace perturbation that CRG also does. In that case, PPFT has strictly fewer privacy protections. This explains why it has worse privacy in the experiments, but why does it have worse utility?
4 comments

Thank you again for your thoughtful review. We would like to clarify our rationale for retaining input perturbation.

In Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation, Yoon et al. directly compare reconstruction performance with and without noise. Figure 5 shows that adding noise reduces reconstruction ROUGE-L across the evaluated pooling sizes, with the largest reduction at (k=4). These results demonstrate that transmitting pooled embeddings without noise can leave substantial information recoverable under the evaluated PPFT attack.

CRG’s positional references do not eliminate information already present in the transmitted embeddings. We therefore retain input perturbation as an additional measure against input reconstruction, while reference prediction and client-side resolution address direct disclosure through generated outputs. We cite the PPFT findings as motivation for this design choice; their measured effect should be understood within the PPFT setting.

We hope this clarification, together with our earlier responses on automatic reference conversion and the non-private baseline, addresses your concerns. If so, we would be grateful if you would reconsider your assessment. We also welcome any remaining questions or points requiring further clarification.

Thank you again for your review. We clarified that CRG encodes the entire original input without domain specific sensitive span selection, and added a task finetuned plaintext baseline distinct from noise free CRG. We also explained how position prediction and client side resolution improve exact reuse compared with reconstructing strings from noisy representations.

We believe these additions directly address the input processing, non private baseline, and utility questions in your review. Could you confirm whether these points are now resolved and revisit your assessment, or indicate which remaining concern is decisive?

Thanks for your response. The new results and clarifications have addressed my concerns. I'm happy to raise the review score accordingly.

AuthorsOpenReview ↗

Thank you for your reconsideration

Thank you for revisiting our paper and for confirming that the additional results and clarifications have addressed your concerns. We appreciate the time you took to reconsider your assessment and your constructive feedback, which helped us clarify CRG’s input processing and strengthen the experimental comparisons.

Reviewer_i7hCOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Client-Resolved Generation (CRG), which separates server-side generation from the realization of input-derived strings. The client transmits pooled and perturbed representations and retains a private request-local position table. The server generates a mixture of ordinary words and positional references, which the client resolves after receiving the response. Public alignment and private domain adaptation train this interface without giving the server the plaintext reference dictionary. Experiments cover medical and document-grounded QA, identifier transfer, and tool calling, with reconstruction, output-inference, raw-logit, and prompt-injection analyses.

Critical strengths
  1. The paper identifies an output channel that input protection alone does not remove. Sensitive values can reappear in server-visible answers even when the input is transmitted through protected representations. CRG trains the server to predict request-local references directly and leaves lexical realization to the client, rather than first generating private strings on the server and masking them afterward. This is a coherent interface-level response to the problem, and the server must learn to locate source positions without receiving plaintext source tokens (Sections 3.1-3.5).
  2. The improvements on tasks requiring exact reuse are substantial. Relative to PPFT, complete-call exact match rises from 57.32% to 79.92% on SealTools, from 32.70% to 66.80% on PIIFORM, and from 74.13% to 92.00% on DOCPII (Table 2). The relative gains increase with reference coverage on SealTools and PIIFORM (Figure 3), which is consistent with the proposed mechanism: selecting an address avoids reconstructing an exact identifier or phrase from perturbed representations. These results provide a concrete utility benefit, even though the advantage is not uniform across all tasks.
Critical weaknesses
  1. The comparison is narrow. Tables 1-2 include only PPFT and AlienLM as competing learned methods; noise-free CRG is a proposed-method configuration. Related Work discusses HaS [1], PrivacyRestore [2], and LatticeGen [3], covering local restoration and output concealment. Their protection targets and server-access assumptions differ, so direct ranking may be inappropriate. Nevertheless, representative comparisons under explicit observation models, or substantive incompatibility explanations, are needed to establish the broader advance.
  2. A simple client-side restoration control is missing: replace private input strings with request-local placeholders, have the server preserve them, then restore strings locally. This may reveal more surrounding semantics than CRG, so equal privacy cannot be assumed; compare utility, field-recovery risk, and cost on the same protected fields. A PPFT variant with the same reference-output vocabulary would further isolate reference prediction from alignment and training changes. Without these controls, the exact-reuse gains do not establish how much the full architecture contributes beyond returning addresses instead of values.
  3. Protection covers reference-addressable lexical content, limiting its privacy value and novelty. Ordinary generated words, new sensitive conclusions, and reference structure remain visible; smaller addressable windows can also return source values to plaintext (Appendices A, C, N). Hiding copied identifiers does not hide a newly generated diagnosis. This is an acknowledged scope limit, but broader significance requires clearer positioning against restoration/pointer interfaces and evaluation of residual semantic exposure, with claims confined to private lexical reuse.
Questions for authors
  1. Could you expand the comparison beyond PPFT and AlienLM to representative local anonymization/restoration or output-concealment methods already discussed in the paper, while explicitly aligning the observation model, protected content, and allowed server modifications?
  2. Could you add a request-local placeholder-and-restoration baseline and a reference-output PPFT control? Please compare utility and actual leakage jointly, rather than assuming that equal noise parameters or the same protected-field list imply equal privacy.
  3. Which sensitive outputs are intentionally left unprotected, particularly values or conclusions absent from the input? Please separate direct lexical exposure, hidden-value recovery, and semantic inference in the claims and evaluation.
5 comments

Thank you for your thorough review and for the positive evaluation of CRG's output-protection problem formulation and exact string reuse performance. Below, we address your comments regarding the scope of comparison baselines, component-level contributions, and protection scope.

1. Adding comparisons with client-resolved restoration methods and clarifying observation conditions.

As requested, we conducted comparisons against a request-local placeholder method with client-side restoration as well as the HaS approach using a learned Seek model. The table below reports the performance of the final client-resolved responses:

MethodSealTools EMPIIFORM EMDOCPII F1
NER placeholder + table restoration84.952.192.0
HaS-max + learned Seek56.138.079.6
CRG, pool2·$\epsilon$7579.966.893.6

The NER method replaces detected strings with request-local placeholders and restores them using a client-side table, whereas HaS-max masks a broader range of spans and reconstructs the response via a learned Seek model. Both baselines in the table allow server-side Stage 2 training on masked inputs, where the server directly observes the unmasked context alongside placeholders. In contrast, CRG transmits the entire private query as pooled and perturbed continuous representations, and the client locally resolves the positional references predicted by the server.

CRG outperforms HaS-max + learned Seek across all three benchmarks and surpasses the simple placeholder baseline on PIIFORM and DOCPII, while trailing the placeholder baseline on SealTools. Masking scope also plays a critical role: for instance, simple NER leaves a substantial portion of clinical fields exposed in DDx-Plus, whereas HaS-max removes direct input-level string exposure across the four evaluated sensitive fields. Accordingly, we will report utility alongside the actual masking scope and explicitly note that this table should not be interpreted as a ranking under equivalent privacy levels.

PrivacyRestore learns type-specific restoration vectors for predefined sensitive span categories and modulates internal server activations to utilize the removed information. By contrast, CRG learns request-local positional references rather than entity-type restoration vectors and performs final string realization exclusively on the client (PrivacyRestore).

Additionally, Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation reports a Pri-NLICE accuracy of 42.77% for 1B PrivacyRestore, whereas CRG and PPFT in our study achieve approximately 53% and 58%, respectively. We will include these figures in the Appendix as cross-literature reference points while explicitly noting differences in training, noise, and scoring protocols. Similarly, we will report related privacy results while clearly distinguishing each method's server-visible signals and attack assumptions.

LatticeGen is a collaborative protocol in which the client controls sampling and conceals the true generation trajectory within a noise-token lattice. Because its generation and communication protocol differs fundamentally from CRG where the server generates a reference-bearing response that is subsequently resolved by the client we will explicitly clarify this distinction rather than presenting it as a baseline evaluated under an identical interface.

2. Analyzing the role of training design via a simple restoration control and a PPFT + reference experiment.

The NER placeholder + table restoration baseline in the first table corresponds to the requested simple client-side restoration control. In addition, we conducted an experiment introducing the exact same 4,352 reference tokens and output codec into a completed PPFT Stage 1 checkpoint during Stage 2, trained and evaluated under identical conditions on the same training data:

MetricPPFT + referenceCRGDifference
SealTools EM33.8979.92+46.03 pp
ToolACE EM9.9446.59+36.65 pp
PolicyQA F131.0342.96+11.93 pp
MultiSpanQA overlap F141.3272.79+31.47 pp

While this control undergoes standard PPFT Stage 1 training, it lacks Stage 1 reference-aware alignment; reference generation is introduced for the first time during Stage 2. Under the same training budget, adding the reference interface post-hoc fails to reproduce CRG's performance.

Notably, on SealTools, PPFT + reference achieved 100% validity in both output formatting and reference addressing, yet argument EM remained at 34.31%. This demonstrates that the bottleneck is not merely an inability to emit valid reference tokens, but rather the ability to select the appropriate input positions from protected representations and bind them to the response. These results support that CRG's contribution extends beyond simply adding address tokens to the output vocabulary, relying instead on joint reference-aware alignment and domain adaptation to master this interface.

We note that this comparison serves as a practical control under a fixed 3-epoch budget and does not imply that post-hoc adaptation would fail under longer training schedules or alternative initializations. Rather than attributing the gap solely to insufficient training steps, we will explicitly specify the differences in both reference introduction timing and training history.

We also measured client-side resolution overhead separately. Across 1,549 DDx-Plus queries, CRG's resolution step including position-table construction and error handling averaged 0.0776 ms (p95: 0.1036 ms) on an EPYC 7513 CPU. Because this measures only the resolution operation itself (excluding encoding, communication, and generation), we will not interpret it as end-to-end latency compared on identical hardware against other methods.

3. Supporting exact string reuse while distinguishing direct exposure from residual inference risks.

The core problem CRG addresses is output-side exposure that arises when a service must reuse exact input strings in its responses. Instead of requiring the server to regenerate those strings in plaintext, CRG predicts positional references and lets the client materialize the original text—thereby decoupling exact lexical utilization on the client from direct plaintext exposure on the server's output path.

We evaluated this effect separately across direct exposure, hidden-value recovery, and semantic inference. On DDx-Plus, direct exposure rates under exact matching of protected strings were 0.32% for sex, 0.59% for symptoms, and 0.73% for medical history.

We also separately evaluated information that does not appear literally in the server output. Under the noise-trained GPT-2 Attack B which jointly observes both the input representation and the unresolved server output exact recovery of hidden sensitive values was 10.62%. Under a learned output-only attack, symptom and medical-history macro-F1 scores were 11.63 and 9.08, compared to 8.51 and 4.07 for a frequency baseline that does not observe the output. Accordingly, we do not equate reduced direct string exposure with the absence of residual inference risk, and we will explicitly report the observation signals and metrics for each evaluation.

Finally, we clarify the protection scope of our approach. Novel diagnoses absent from the input, general explanatory text, the structural pattern of references, and plaintext values generated outside the reference scope are not concealed by this interface alone. CRG's contribution lies not in hiding all semantic information, but in reducing direct server-side plaintext realization in tasks requiring exact string reuse. In the revised manuscript, we will explicitly annotate this protection scope alongside the observation conditions for each attack in our result tables.


Thank you again for your time and insightful feedback. Please let us know if you have any further questions or comments we would be glad to address them during the discussion period.

Thank you again for your detailed feedback. We have added both requested restoration baselines and the PPFT reference output control. Under the fixed Stage 2 budget, the latter achieves 33.89% SealTools EM versus 79.92% for CRG, directly addressing whether adding reference tokens alone reproduces the gains. We also report residual inference and explicitly distinguish lexical exposure reduction from semantic confidentiality.

We believe these additions address the main comparison and contribution isolation concerns. Could you confirm whether these concerns are now resolved and reconsider your assessment accordingly, or clarify the specific issue that remains?

Reviewer_u4c6OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PPFT-style text-free inference. The client encodes the input, mean-pools pairs of tokens, adds noise, and only transmits these vectors. Then a Llama-3.2-1B decoder at the server side produces the answer. Instead of copying words from input as plaintext, the decoder produces position tokens, and the client recovers words at the position tokens. There are two phases of training: aligning tokens on public data, then LoRA tuning on noised private data. The results demonstrate that, for copying-centric tasks like Seal Tools exact match, the scores increase from 57.3 to 79.9 over PPFT.

Critical strengths
  1. This paper is well motivated. Protecting only the input does not help if the output reproduces the private content. We make this concrete with the prompt-injection results.

  2. The mechanism is simple, applies on top of text-free inference, and preserves the privacy of the server model.

  3. The line-coverage analysis reveals exactly when the technique is useful and when it is not.

  4. This privacy assessment is more extensive than one usually finds for this type of work. They also notice the "leakage" that persists such as the +10 point age lift.

Critical weaknesses
  1. No guaranteed guarantee of privacy and ε = 75 is meaningless. In 1024-dimensional vectors, ε = 75 under metric DP provides no privacy protection. All privacy claims are empirical and rely on the particular attackers used.

  2. The attackers are feeble. Attacks A and B are GPT-2-medium trained on ~5.9K clean embeddings. The paper then blames clean-to-noisy shift for poor recovery, but a real attacker would have trained on noised embeddings as the noise mechanism is public. Attack C runs off-the-shelf LLMs with no information about the data distribution.

  3. Unresolved output continues to leak a great deal. The text of explanatory sentence and in many cases diagnosis remain in plain text e.g., "suggests meningeal irritation" in the examples. The position of reference tokens and the length of the output will also indicate structure. On templated data e.g. DDx-Plus a reference position can reveal which field it is referencing. That all goes unquantified.

  4. Novelty is limited. The method is a combination of PPFT+Pointer/copy. Placeholder-and-restore schemes like HaS are essentially keeping the mapping on the client. The twist is the predicted locations from pooled, noised vectors, but the authors should say this more explicitly. HaS and Privacy Restore are mentioned without experimental comparison.

Questions for authors
  1. A and B are training on noised embeddings instead of clean inputs? How much do the input and hidden-value recovery rates increase?

  2. Can a malicious server train a probe at reference positions on the decoder hidden states in order to predict the referenced word? This feels like the most straightforward attack to the output-side claim.

  3. What is the experiment exactly: how does CRG compare to HaS or Privacy Restore on the "shared structure of "resolve on the client"" for the PII and tool-calling tasks?

  4. What is the non-private upper bound i.e. the same decoder but with plaintext input (ref Tables 1-2)? How stable are results across seeds?

4 comments

Thank you for your thoughtful review and for recognizing the importance of output-side privacy. Below, we address your comments regarding attacker adaptability, internal representation exposure, comparisons with related methods, and un-protected performance baselines.

1. Additional evaluation of noise-adapted Attacks A and B

We agree that an adversary should be assumed capable of adapting to the public perturbation mechanism. Maintaining the exact training budget and evaluation protocol of our GPT-2-medium reconstructor, we applied ϵ=75\epsilon=75 Laplace noise to the training embeddings as well:

Attack / MetricClean Train →\rightarrow ϵ=75\epsilon=75 EvalNoisy (ϵ=75\epsilon=75) Train →\rightarrow ϵ=75\epsilon=75 Eval
Attack A: Input reconstruction (ROUGE-L)36.4729.11
Attack B: Hidden sensitive value exact recovery10.56%10.62%

For Attack A, noisy training did not improve reconstruction performance across any of the three training seeds. For Attack B, the observed difference between clean and noisy training was also negligible. Thus, within the tested GPT-2 attacker configuration, eliminating the clean-to-noisy distribution mismatch alone does not improve reconstruction performance. Specifically, under noisy training, the attacker's generations frequently degraded to the point of failing to recover even the underlying dataset templates, while sensitive attribute recovery remained minimal and comparable to the clean-trained baseline.

2. Additional analyses of reference-output suppression and cross-dataset hidden-state attacks

Thank you for pointing out the potential risk of recovering sensitive strings from internal representations. To examine this, we conducted two distinct analyses: an inference-time diagnostic that directly suppresses the reference output pathway, and a reconstruction attack trained on auxiliary medical data.

First, in Appendix L.5, we kept the served CRG model, DDx-Plus inputs, and ϵ=75\epsilon=75 perturbation fixed while blocking the generation of all 4,352 reference tokens during decoding. Across all six prompt-injection conditions without additional fine-tuning, sequence-level ROUGE-L ranged from 0.03 to 1.97, and the recovery lift over the shuffled-record control was at most 1.28 and 0.59 percentage points for symptoms and medical history, respectively. Thus, merely prohibiting reference tokens does not cause the decoder to revert to faithful plaintext reconstruction of the patient description.

In a separate raw-logit diagnostic, the Hit@10 rate of the corresponding single-token plaintext at reference generation steps was 2.86% on DDx-Plus and 0.27% on NLICE.

Additionally, we evaluated hidden-state reconstruction under a setting where the attacker has no access to training examples from the target dataset. Freezing the CRG (k=2,ϵ=75k=2, \epsilon=75) model, we trained a full-vocabulary reconstruction head exclusively on auxiliary medical data and evaluated it on the target set:

Overall Sensitive WordsSymptom WordsHistory WordsSex Words
6.83%8.18%2.95%47.35%

All values represent word-level exact match rates across referenced words belonging to each field, rather than full-field recovery rates per patient. These results indicate that attaching a vocabulary prediction head alone is insufficient to recover the original clinical text. We will include these analyses in the Appendix for the camera-ready revision.

3. Clarifying CRG's distinct contributions and comparisons with restoration-based methods

CRG builds upon PPFT's continuous input-protection pathway. Its core technical challenge lies in training the server to generate request-local positional references directly from pooled and perturbed representations without receiving raw source tokens or client-side position tables. Our Stage 1 reference-aware alignment and Stage 2 domain adaptation enable exact lexical reuse under this protected interface.

Regarding related restoration-based approaches, HaS masks sensitive spans in the input and employs a learned client-side Seek model for de-anonymization, so characterizing it purely as a rule-based method would be inaccurate. PrivacyRestore learns type-specific restoration vectors for predefined privacy spans and conditions internal server activations to incorporate the removed information. In contrast, rather than pre-selecting sensitive spans or learning entity-type vectors, CRG transmits the entire private query as continuous representations, predicts source positions for input-matching output words, and defers final string realization entirely to the client. We will clearly articulate these differences in preparation pipelines and restoration mechanisms in the revised manuscript.

Below are the results from our completed comparisons against restoration-based baselines:

MethodSealTools EMPIIFORM EMDOCPII F1
NER placeholder + table restoration84.952.192.0
HaS-max + learned Seek56.138.079.6
CRG (k=2,ϵ=75k=2, \epsilon=75)79.966.893.6

CRG outperforms HaS-max + learned Seek across all three benchmarks and surpasses the NER placeholder baseline on PIIFORM and DOCPII, while trailing the placeholder baseline on SealTools. Because masking scopes and server-visible signals differ substantially across these methods, we will explicitly note that these results should not be interpreted as direct rankings under equivalent privacy guarantees.

Additionally, in Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation (Yoon et al., 2026), the 1B model achieves an accuracy of 0.43 on NLICE, which is lower than both CRG (0.53) and PPFT (0.58) under equivalent noise settings. We will include these comparison metrics alongside their corresponding privacy exposure levels in the Appendix.

4. Adding un-protected Plaintext LLM baselines and multi-seed stability results

Because noise-free CRG still operates through the reference-based generation interface, we explicitly distinguish it from a standard un-protected baseline. Below, we compare CRG against a task-finetuned Plaintext LLM built on the same Llama-3.2-1B decoder backbone:

ModelSealTools EMPIIFORM EMDOCPII EMFlashcards Lexical F1
Plaintext LLM, task-finetuned92.0587.8798.7346.25
CRG no-noise87.0376.2395.4044.45
CRG (k=2,ϵ=75k=2, \epsilon=75)79.9266.8092.0042.70

This comparison illustrates the task-specific utility trade-offs introduced by the privacy interface. Since the models undergo different training pipelines, we will note that the observed gap should not be attributed solely to reference conversion.

Finally, evaluating across three random seeds (42, 43, and 44) confirms consistent stability, achieving 66.82±0.2666.82 \pm 0.26 on PIIFORM, 92.23±0.2092.23 \pm 0.20 on DOCPII, and 42.68±0.0542.68 \pm 0.05 on Flashcards.


Thank you again for your time and constructive feedback. Please let us know if you have any further questions or concerns we would be happy to discuss them further.

Thank you again for your review. We have added noise-adapted reconstruction attacks, an auxiliary data trained hidden state reconstruction head, restoration baselines, and a same-decoder plaintext reference. Within the tested attacker configuration, noise adapted training changed hidden value recovery only from 10.56% to 10.62%. We have also clarified the empirical scope and explicitly reported residual leakage.

We believe these additions substantially address the experimental concerns raised in your review. Could you revisit your assessment in light of these results and indicate which specific concern, if any, remains decisive?

AuthorsOpenReview ↗

Additional validation and key results addressing your main concerns

We summarize the additional experiments conducted in response to your main concerns and their findings below.

  • We directly evaluated noise-adapted attackers. Using the same attacker and training budget, noise-adapted training changed Attack A ROUGE-L from 36.47 to 29.11 and Attack B hidden-value recovery from 10.56% to 10.62%. Within the evaluated configuration, removing the training–evaluation noise mismatch did not substantially improve reconstruction performance. These results directly address the concern that the previous findings were attributable to this mismatch.
  • We added learned reconstruction attacks on server-side hidden states. Beyond the existing raw-logit analysis, we trained a full-vocabulary reconstruction head on auxiliary medical data to recover plaintext from hidden states. Overall sensitive-word exact match under this transfer attack was 6.83%. This extends the evaluation beyond visible output strings to attacks that exploit the server’s internal representations.
  • We directly compared CRG with closely related restoration methods. PIIFORM EM was 66.8% for CRG, 52.1% for NER placeholders, and 38.0% for HaS-max with learned Seek. These results demonstrate concrete exact-reuse gains on this benchmark over closely related approaches that also perform client-side restoration.
  • We added a non-private reference and evaluations across seeds. We provided a task-finetuned plaintext-input model using the same decoder and reported evaluations with seeds 42, 43, and 44. These additions distinguish noise-free CRG from a standard plaintext model and strengthen the comparison beyond a single run.

These additions provide concrete experimental evidence addressing your questions about attacker adaptation, internal-state recovery, closely related baselines, and the non-private reference. We therefore ask that you reconsider your final assessment in light of this evidence. If any concern remains decisive, please identify the specific issue so that we can address it directly.

Comments on the paper

2 comments

Thank you for requesting a direct comparison of privacy exposure alongside utility. We conducted an additional LLM-as-judge evaluation on 650 NLICE cases and 3,000 DOCPII cases, comparing CRG’s unresolved server output against PPFT’s output, HaS-max’s output, and HaS-max’s sanitized input together with its output.

We used GPT-6.1-sol to compare the supplied server-visible texts from the same test case and identify which revealed more sensitive information. References and placeholders were treated as hidden values, without access to the original private input or restoration mappings. Method names were withheld, and each pair was presented in both orders. Consistent preferences were reported as wins; all remaining cases were grouped as ties.

DatasetComparator observationCRG less exposedComparator less exposedTie
NLICEPPFT output5981933
NLICEHaS-max output421126103
NLICEHaS-max input + output406130114
DOCPIIPPFT output1,49801,502
DOCPIIHaS-max output3232,965
DOCPIIHaS-max input + output15862,836

DOCPII shows that CRG combines accurate private-value reuse with exposure results predominantly tied with the extended masking baseline. CRG reduced exposure relative to PPFT, while most comparisons with HaS-max resulted in ties—even when HaS-max’s sanitized input was included. Combined with CRG’s higher client-resolved utility reported previously, this demonstrates a favorable privacy–utility result: improved task performance without a systematic increase in judged exposure of structured personal information.

NLICE demonstrates an additional advantage for clinical information that remains exposed under HaS-max. CRG was more frequently judged less exposed both in output-only comparisons and when the remaining input context visible to the HaS server was included. The advantage therefore extends beyond structured identifiers to a clinical setting, where sensitive information is expressed through patient descriptions and contextual statements. This is consistent with CRG’s design: input-matching clinical expressions can be generated as client-resolved references without first requiring a sensitive-entity detector to recognize them.

These results support CRG’s ability to protect input-derived personal information while additionally reducing clinical information exposure that persists under HaS-max. Together with the utility improvements, they demonstrate the practical benefit of learning to reuse private content through references while leaving its plaintext realization to the client.


Thank you again for your thoughtful feedback. We would appreciate your consideration of the additional experiments and clarifications, and your reassessment if these address your concerns. Please let us know if any concerns remain or further clarification would be helpful.

Thank you again for your review. As the discussion period ends on November 18 at 23:59 AoE, we would appreciate knowing whether our responses have addressed your main concerns or whether any questions remain. If the clarifications and additional results have addressed your concerns, we would be grateful if you could consider updating your assessment. We would be happy to clarify any remaining points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.