acceptodds
Under review as a conference paper at ICLR 2027

DP-MIND: Differentially private memory-guided Long context Inference

Abstract

LLM-as-a-Service has become a dominant paradigm for accessing large language models without requiring local deployment. However, the process raises severe privacy concerns when user queries may include sensitive or confidential contexts. Differential privacy (DP) provides a principled approach to protecting such inputs, but existing DP inference methods are primarily designed for short- or moderate-context settings and struggle with million-token-scale inputs due to limited context capacity and severe utility degradation under DP perturbation. In this paper, we propose DP-MIND, a memory-guided framework that scales differentially private inference to million-token contexts. DP-MIND enables private long-context inference by locally perturbing the input and processing it chunk by chunk with a compact memory that preserves task-relevant information across the sequence. To make this memory robust to DP noise, we optimize the memory construction policy directly on perturbed inputs using reinforcement learning, guided by both final answer correctness and memory consistency. We further recover perturbation-corrupted answer spans through a local denoising module without exposing the original context to the external model. We provide theoretical privacy analysis of DP-MIND and further conduct extensive empirical analysis across multiple long-context benchmarks with context lengths up to the million-token scale, demonstrating that DP-MIND consistently outperforms existing DP and long-context inference baselines in both scalability and utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential Privacy, LLM Inference, Long Context Inference

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_9VZ1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DP-MIND, a framework for differentially private long-context LLM inference. DP-MIND perturbs the input locally and process it chunk by chunk with a memory. They optimize the memory construction policy on the perturbed inputs using reinforcement learning. They perform experiments on RULER-HQA and NIAH with context lengths up to the million-token scale, and show that DP-MIND can outperform baseline methods.

Critical strengths
  • The paper focuses on differentially private inference over long contexts, which is an interesting problem.

  • The paper is easy to follow.

  • The experiments cover context lengths up to 1.8 million tokens, showing the scalability of the proposed framework.

Critical weaknesses
  • While DP inference to long contexts is an interesting problem, it is unclear to me why existing RL-based memory methods (e.g., MemAgent) are insufficient under DP perturbation.

  • The proposed framework mainly combines existing DP perturbation and RL-based memory inference methods. Could the authors further clarify the technical novelty of the proposed approach?

  • The experiments are limited to the Qwen2.5 model family. I recommend evaluating the proposed method on more model families to show its generalizability.

  • It would be better to perform ablation studies on the memory-consistency reward RmemR_{mem}. I recommend comparing the performance with and without this reward.

  • I recommend including existing work on differentially private in-context learning [1, 2, 3] in the literature review and discussing its relevance to the proposed approach.

[1] Wu, Tong, et al. "Privacy-preserving in-context learning for large language models." International Conference on Learning Representations. 2024.

[2] Tang, Xinyu, et al. "Privacy-preserving in-context learning with differentially private few-shot generation." International conference on learning representations. Vol. 2024. 2024.

[3] Gao, Fengyu, et al. "Data-adaptive differentially private prompt synthesis for in-context learning." International Conference on Learning Representations. 2025.

Questions for authors

See Weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 9VZ1 (1/2)

We thank the reviewer for the detailed review and valuable suggestions. The reviewer raised important concerns about the motivation for DP-aware memory adaptation, technical novelty, ablations, and generalizability. We address each concern below.

W1: Insufficient justification for adapting existing memory methods to DP

We apologize for the confusion regarding the motivation for DP-aware memory adaptation. DP perturbed entities and relations can be stored in memory and persist through subsequent updates, causing errors to accumulate. Existing RL-based memory methods such as MemAgent are designed for memory construction on clean inputs, without explicitly addressing this noise-induced error accumulation. Consequently, directly applying them to DP-perturbed inputs provides insufficient robustness.

Figure 2 and Appendix Table 3 compare RL MemoryAgent, DP-RL MemoryAgent, and the full DP-MIND on RULER-HQA at ϵ=7\epsilon=7. At 448K, DP-aware training improves SubEM from 67.19 to 71.88, with no further gain from local answer denoising. At 56K, training leaves the score unchanged at 67.97, while local denoising increases it to 73.44. These results show that RL-based memory inference alone does not fully address the utility loss introduced by DP perturbation,


W2: Technical novelty beyond combining existing methods

DP-MIND goes beyond combining an existing perturbation mechanism with a memory agent by introducing DP-aware optimization of memory construction together with local answer denoising.

  • Memory-consistency reward for robust memory construction. Final-answer correctness provides a terminal reward, but does not directly measure whether intermediate memories retain the supporting evidence. We therefore train a reference policy on clean data to generate memory trajectories. During RL on DP-perturbed inputs, the memory-consistency reward encourages denoised memories to match their clean references by penalizing the average embedding distance between them. Denoising reduces penalties for recoverable token substitutions. We combine this reward with correctness of the final locally denoised answer, supervising both intermediate information retention and final-answer utility.

  • Local answer recovery. The user retains the clean context and token alignment locally and uses them to restore recoverable perturbed answer spans. This improves answer utility while keeping restored sensitive information on the user's device.

  • Privacy analysis and evaluation at the million-token scale. We provide formal privacy guarantee of the whole framework and evaluate inference up to 1.8M tokens. We build on established DP and memory primitives; the contribution lies in adapting memory training and answer recovery to this setting.

We will clarify these contributions and add the clean-reference training details to the revised manuscript.


W3: Limited evaluation across model families

To evaluate whether the observed behavior generalizes beyond the Qwen2.5 family, we have conducted preliminary experiments using Llama-3-8B-Instruct (8K context window) and Llama-3-8B-Instruct-Gradient-1048k (approximately 1M context window), which share the Llama 3 base model. The results below cover direct inference, memory-augmented inference without RL, and the extended-context model. To ensure a fair comparison, all methods are evaluated on the same set of pre-generated perturbed inputs (ϵ=7\epsilon=7) used in Qwen experiments, with performance measured by Sub-EM (%). Due to the substantial time required for RL training, we have not yet completed the full evaluation.

Method7K14K28K56K112K224K448K
Llama-3-8B-Instruct44.53------------------------------
Memory-augmented Llama-3-8B-Instruct (without RL)44.5336.7235.9422.6625.0018.7514.06
Llama-3-8B-Instruct-Gradient-1048k35.9437.5032.8134.3827.3422.6614.06

We will include these results and complete the remaining RL experiments in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 9VZ1 (2/2)

W4: Missing ablation on the memory-consistency reward

As you suggested, we performed the requested ablation to isolate the contribution of the memory-consistency reward beyond the final-answer reward. The reference memory is generated by a policy first trained on clean inputs for task performance. We encourage the denoised memory from perturbed inputs to approach this reference, providing intermediate supervision for preserving task-relevant information under DP noise.

We compare DP-RL MemoryAgent with and without the memory-consistency reward at context lengths from 7K to 448K, alongside the original RL MemoryAgent baseline. The results are reported below.

Method7K14K28K56K112K224K448K
RL MemoryAgent72.6667.9767.1967.9765.6269.5367.19
DP-RL MemoryAgent w/o memory-consistency reward73.4469.5369.5367.1966.4169.5370.31
DP-RL MemoryAgent w/ memory-consistency reward75.0071.8869.5367.9766.4169.5371.88

Adding the memory-consistency reward improves SubEM with gains of up to 2.34 percentage points, supporting its benefit beyond the final-answer reward.


W5: Missing discussion of differentially private in-context learning

We will include all three papers in the related work and discuss their connection to DP-MIND. These works enable privacy-preserving LLM inference by protecting in-context examples through differentially private response aggregation or synthetic demonstration generation, while DP-MIND focuses on protecting long context LLM inference.

Thank you & Looking forward to your feedback

We would like to extend heartfelt thanks to you for your time and efforts throughout the review process, and for your positive assessment of our work. Each of your comments has led to a concrete change in the paper. For your convenience, we briefly recap them here:

  • W1, necessity of DP-aware memory adaptation. We clarified that memory methods trained on clean inputs do not address the accumulation of perturbed entities and relations across memory updates. At 448K, DP-aware training improves Sub-EM from 67.19 to 71.88, showing that directly applying an existing RL memory method is insufficient under DP perturbation.

  • W2, technical novelty. We clarified that DP-MIND introduces a memory-consistency reward that supervises intermediate information retention under DP noise, together with local answer denoising that restores recoverable spans without exposing clean information to the server. We also provide an end-to-end privacy analysis and evaluation at context lengths up to 1.8M tokens.

  • W3, generalization across model families. As you suggested, we plan to extend the full evaluation to the Llama family, including RL training. As a first step, we conducted preliminary experiments with Llama-3-8B-Instruct and Llama-3-8B-Instruct-Gradient-1048k using the same perturbed inputs as in our Qwen experiments, thereby extending the current evaluation beyond the Qwen2.5 family.

  • W4, memory-consistency reward ablation. We added the requested comparison between DP-RL MemoryAgent with and without the memory-consistency reward over context lengths from 7K to 448K. The reward improves Sub-EM by up to 2.34 percentage points beyond training with the final-answer reward alone.

  • W5, differentially private in-context learning. We added the three suggested papers and clarified their relationship to DP-MIND. These methods protect in-context examples through private aggregation or synthetic demonstration generation, whereas DP-MIND protects long contexts during LLM inference.

We hope these revisions address your concerns and clarify the motivation, technical contributions, and generalizability of DP-MIND. We would be happy to provide any further clarification.

Reviewer_FT5oOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a memory guided framework for private long context inference. The paper divides the long context into chunks and only shares perturbed chunks to the external LLM. Locally the original clean chunks are kept. The sequential response of the LLM on each perturbed chunk along with updated memory from previous responses is used to guide the next response of the LLM. This per-step updated memory and sequential responses, gives a DP output at the final step. Post this, a denoising algorithm cleans the generated DP answer which uses the alignment between local clean context and perturbed context. The memory construction policy is updated using reinforcement learning techniques to preserve task accuracy and memory consistency.

The experiments demonstrate that DP-MIND performs better than baselines for large contexts and scales well. Ablations demonstrate the effects of chunk size on accuracy and efficiency and highlight the gains from the final denoising stage.

Critical strengths
  1. Clean segregation between local processing and information shared with external LLM which helps to keep the privacy analysis simple and elegant.
  2. Memory construction policy designed to improve both downstream task accuracy and consistency with clean chunks of data.
  3. Denoising algorithm records token level alignment of perturbed and clean chunks and uses this to clean final DP answer locally.
Critical weaknesses
  1. It is not clear what the metric DP guarantees here. Since epsilon in metric DP depends on the embedding geometry, its meaning is not directly comparable to standard DP.
  2. How the reinforcement learning fits in the final external LLM inference process is unclear to me. More explanation on who trains the memory construction policy (user or external LLM) in the deployment setting would be helpful.
Questions for authors
  1. Could the authors elaborate on which embedding table is used for the perturbation across all the experiments?
  2. Could the authors report the fraction of tokens changed by the perturbation at each epsilon, and typical distances between neighbouring token embeddings?
  3. Could the authors state the guarantee for a sensitive span of k tokens (e.g., a name) and discuss repeated occurrences?
4 comments
AuthorsOpenReview ↗

Response to Reviewer FT5o (1/2)

We thank the reviewer for the constructive comments and valuable suggestions. We address the concerns as follows.

W1: Interpretation of the metric DP guarantee

We apologize for the confusion regarding the interpretation of our metric-DP guarantee. Metric DP limits how much the released output can distinguish two candidate original tokens. Standard ϵ\epsilon-LDP bounds their output-probability ratio uniformly by eϵe^\epsilon; Definition 3.2 instead gives the distance-dependent bound eϵd(x,x′)e^{\epsilon d(x,x')}, where dd is the Euclidean distance in the fixed embedding space. Closer tokens receive a tighter bound.

For example, at ϵ=7\epsilon=7 and distance 0.8510.851, the sampled median nearest-neighbor distance reported below, the bound is e7×0.851≈386.45e^{7\times0.851}\approx\mathbf{386.45}. This is an output-probability bound, not an attack success probability or a token change rate. A uniform vocabulary-wide bound would require the maximum pairwise distance DD, giving eϵDe^{\epsilon D} per token. The sampled nearest-neighbor median is not a substitute for DD. Thus, ϵ\epsilon must be interpreted together with the embedding geometry and scale, and equal numerical values do not imply equivalent protection under metric DP and standard LDP.

We use metric DP for our main experiments, following its use in text perturbation [1,2] and privacy-preserving LLM inference systems [3,4]. Specifically, we independently perturb Qwen2.5-7B token embeddings and map them back to tokens, using Euclidean embedding distance as the privacy metric.

We evaluate DP-MIND under both metric DP and LDP-based guarantees. Specifically, Table 2 covers metric DP (dχd_\chi-privacy and SanText) and ϵ\epsilon-LDP within a given random adjacency set (RANTEXT from InferDPT), at context lengths from 56K to 448K. DP-MIND consistently outperforms the corresponding Qwen-1M baseline, with gains of 22.66–68.75 percentage points, supporting its effectiveness under both evaluated privacy formulations. We will clarify their respective definitions, assumptions, and privacy parameters in the revised manuscript.

Refenrence

[1] Oluwaseyi Feyisetan et al. Privacy- and Utility-Preserving Textual Analysis via Calibrated Multivariate Perturbations. WSDM, 2020.

[2] Xiang Yue et al. Differential Privacy for Text Analytics via Natural Text Sanitization. Findings of ACL-IJCNLP, 2021.

[3] Jeongho Yoon et al. Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation. ACL, 2026.

[4] Amrita Roy Chowdhury et al. Preempt: Sanitizing Sensitive Prompts for LLMs. NDSS, 2026.


W2: Unclear RL training and deployment workflow

We acknowledge that the training and deployment workflow was not sufficiently explicit in the current manuscript. In our intended deployment, the memory policy is trained offline by the LLM provider, using clean reference memories and RL on perturbed training inputs. The provider then serves the resulting DP-aware checkpoint. Users do not train the policy or perform RL updates during inference.

The user first perturbs the context locally and retains the clean context and token alignment on their device. The external service uses the trained checkpoint and the memory update prompt in Appendix A.3 to update memory from perturbed chunks and previous memory. The final answer prompt requests an answer using only DP-derived information and the fixed question. The user then denoises the returned answer locally. We will make the hosted-checkpoint requirement explicit and unify the final-call description in Section 4.1 and Appendix A.3.


Q1: Embedding table used for perturbation

Across all experiments, we consistently use the pretrained input embedding table of Qwen2.5-7B for token perturbation. Each token is mapped to its embedding through a direct lookup, and the embedding table remains fixed throughout training and inference. We will explicitly specify this in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer FT5o (2/2)

Q2: Empirical strength of perturbation

As you suggested, we report the token change rates on 128 contexts from our RULER-HQA measurement set, comprising 7,187,567 original Qwen tokens. We use token-level LCS alignment between the original and released perturbed contexts to account for changes in token boundaries after decoding and re-tokenization.

ϵ\epsilonOverall token change rate
620.70%
76.01%
81.32%

Separately, we randomly sample 2,048 non-special vocabulary tokens from the Qwen2.5-7B input embedding table and perform an exact nearest-neighbor search over the full non-special vocabulary, excluding each token itself.

StatisticNearest-neighbor L2 distance
Mean0.775
25th percentile0.765
Median0.851
95th percentile0.945

The change rates describe the observed perturbation, while the sampled distances help interpret the embedding metric. Neither replaces the formal privacy guarantee. We will include the tables and measurement procedures in the revised manuscript.


Q3: Privacy guarantees for sensitive spans and repeated occurrences

For sensitive spans and repeated occurrences, we independently apply the token-level mechanism to each position, including repeated occurrences of sensitive information. For a fixed question and two equal-length contexts X,X′X,X' differing in a kk-token sensitive span, composition gives

\leq \exp\left(\epsilon\sum_{i=1}^{k}d(x_i,x'_i)\right) \Pr[T(X')\in A],$$ where $T$ denotes the **complete server-visible transcript**, including perturbed chunks, generated memories, and the pre-denoising answer, for a fixed trained model. If that span occurs $r$ times and every occurrence is replaced by the same alternative span in $X'$, the bound is $$\Pr[T(X)\in A] \leq \exp\left(r\epsilon\sum_{i=1}^{k}d(x_i,x'_i)\right) \Pr[T(X')\in A].$$ For different substitutions at different occurrences, the exponent instead sums the distances over all $rk$ affected positions. The factor $r$ counts changed input occurrences; it does not assume shared perturbations. Unlike standard token-level $\epsilon$-LDP, our metric-DP accounting is **distance-dependent**, so it does not generally reduce to $k\epsilon$ or $rk\epsilon$ without a bound on the token distances. We do not follow the mapping-table approach of DP³ [5] , which **matches the current unperturbed token against stored original tokens to retrieve the corresponding perturbed replacement**. This lookup depends on original token identities and therefore **cannot be justified solely as post-processing** of previously privatized outputs. We instead perturb each occurrence independently and account for privacy loss through composition. **Refenrence** [5] Lele Zheng et al. DP³: Differentially Private Prompt Perturbation for Multi-turn LLM Inference. Findings of ACL, 2026.

Thank you & Looking forward to your feedback

We would like to express our sincere gratitude for your time and effort throughout the review process, as well as for your encouraging and constructive assessment of our work. Your comments and questions have helped us improve the manuscript. For ease of reference, we briefly summarize the resulting revisions below:

  • W1, interpretation of metric DP. We clarified that metric DP bounds output-probability ratios based on token-embedding distances, so its privacy parameter must be interpreted with the embedding geometry and is not directly comparable to standard LDP. We also distinguish the formal guarantee from token change rates and nearest-neighbor distances. Table 2 further evaluates DP-MIND with both metric-DP mechanisms and the LDP-based RANTEXT mechanism, showing consistent improvements over Qwen-1M under the evaluated privacy formulations.

  • W2, RL training and deployment workflow. We clarified that the LLM provider trains the memory policy offline using perturbed training inputs and clean reference memories, and subsequently hosts the resulting DP-aware checkpoint. Users do not conduct RL training or update the policy during inference. In deployment, the user perturbs the context locally and retains the clean context and token alignment; the external service processes only the perturbed chunks using the trained checkpoint; and the user applies answer denoising locally. We will state this division of responsibilities explicitly in Section 4.1 and Appendix A.3.

  • Q1, embedding table used for perturbation. All experiments use the pretrained input embedding table of Qwen2.5-7B. Token embeddings are obtained through direct lookup, and the embedding table remains fixed throughout both training and inference.

  • Q2, empirical perturbation strength. On 128 RULER-HQA contexts containing 7,187,567 tokens, the token change rates are 20.70%, 6.01%, and 1.32% at privacy parameters 6, 7, and 8. Exact searches using 2,048 sampled non-special tokens give a median nearest-neighbor L2 distance of 0.851 and a 95th percentile of 0.945. We will add these results and the measurement procedure.

  • Q3, sensitive spans and repeated occurrences. We added the composition guarantee for a kk-token sensitive span occurring rr times. For identical substitutions, the bound is exp⁡ ⁣(rϵ∑i=1kd(xi,xi′))\exp\!\left(r\epsilon\sum_{i=1}^{k} d(x_i,x'_i)\right); for different substitutions, the exponent sums the distances over all rkrk affected positions. Each occurrence is perturbed independently.

We hope these revisions address your concerns and clarify the privacy guarantee and deployment workflow. We would be happy to provide further clarification if any issue remains unsolved.

AuthorsOpenReview ↗

A final check before the discussion period closes

Dear Reviewer FT5o,

As the discussion period approaches its close, we wanted to check in once more. In response to your comments, we clarified how the metric-DP parameter must be interpreted with the embedding geometry, specified the respective responsibilities of the provider and user during training and deployment, and identified the fixed Qwen2.5-7B embedding table used throughout our experiments. We also added the requested token change rates and nearest-neighbor distance statistics, as well as the composition guarantee for sensitive spans and repeated occurrences.

We would appreciate knowing whether these additions address your concerns or whether any point still requires clarification. We will respond promptly to any further questions during the remaining discussion period.

Thank you again for your thoughtful and constructive review.

Reviewer_4s3jOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Provides a method for differentially private LLM inference, that is claimed to be not only private but also scales to million sized token context windows. The paper perturbs each token in any chunk by generating a corresponding embedding for each token and adds differential privacy inducing noise to that embedding and then back projects the noised embedding to a token which is differentially private. It then does this a token at a time and moves to the next new token, and once a chunk is done it again sweeps through the next chunk instead, as the external LLM only receives the fully DP perturbed chunks, and this way it completes the processing of the privacy sensitive document. So there is a chunk wise meta algorithm which internally calles the token wise DP mechanism.

Critical strengths

The paper is quite relevant to current day LLM inference and privacy issues that surround them. The scalability, and elegance of the method, in terms of simplicity of needed implementation is a plus.

Critical weaknesses

are the methods being compared to non private methods? why not compare them with private llm inference methods listed in the related work in the small context setting, if they dont scale to long context, or ablate to the extent of context size they are scalable to?

minor: there are writing issues, that are important to be addressed. -There is a misplaced ? in the beginning of the citation 'Feyistan', as an un-rendered citation or symbol, between lines 138-139.

Questions for authors
  1. ) Is the embedding from each token obtained via a deep embedding or via matrix E applied as a matrix multiplication? 2.) if the latter then can't the query to be privatized instead just be that matrix multiplication, which would then be a linear query, whose workloads have been well studied in DP literature.
3 comments

We thank the reviewer for the valuable feedback on baseline comparisons, citation issues, and the embedding mechanism's connection to linear queries. We address each concern below.

W1: Insufficient comparison with private inference and long-context methods

  • 1. DP LLM Inference Methods

As suggested, we compare direct inference, InferDPT, and DP-MIND on Qwen2.5-7B using identical DP-perturbed inputs, ensuring the same privacy budget (ϵ=7\epsilon=7) and underlying privacy guarantee across methods. For InferDPT, we also enable its local extraction stage using Qwen2.5-7B and adapt it to QA by extracting the final answer from the remote model's response.

Method7K14K28K56K
Direct inference47.6646.0940.62---
InferDPT[1]50.0057.8145.31---
DP-MIND75.7874.2271.0973.44

Although InferDPT improves over direct inference at 7K, 14K, and 28K, DP-MIND achieves further gains of 25.78, 16.41, and 25.78 percentage points, respectively. At the length of 56K, the input exceeds the supported context length of Qwen2.5-7B for both direct inference and InferDPT, whereas DP-MIND supports this length through sequential memory processing and achieves 73.44 SubEM. These results demonstrate the benefit of our framework for inference over longer DP-perturbed contexts.

  • 2. Non-private Long Context Inference Methods

We apologize for the confusion regarding our comparisons with non-private long-context methods. Specifically, Table 1 includes Longformer-style sliding-window attention [2] and direct chunk-wise inference with recurrent memory. The RL-based MemAgent comparison [3] appears in Figure 2 and Appendix Table 3. As you suggested, we additionally evaluate FlexPrefill [4], XAttention [5], and KVzip [6] under the same perturbed inputs (ϵ=7\epsilon=7). All methods use Qwen2.5-7B-Instruct and are evaluated on the same HotpotQA instances using Sub-EM (%).

Method28K56K112K224K
Direct inference40.62---------
FlexPrefill19.5315.627.810.00
XAttention39.8421.882.340.00
KVzip40.6319.534.690.00
DP-MIND71.0973.4468.7570.31

References

[1] Meng Tong et al. InferDPT: Privacy-Preserving Inference for Closed-Box Large Language Models. IEEE Transactions on Dependable and Secure Computing, 2025.

[2] Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The Long-Document Transformer. arXiv:2004.05150, 2020.

[3] Hongli Yu et al. MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-Based Memory Agent. ICLR, 2026.

[4] Xunhao Lai et al. FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference. ICLR, 2025.

[5] Ruyi Xu et al. XAttention: Block Sparse Attention with Antidiagonal Scoring. ICML, 2025.

[6] Jang-Hyun Kim et al. KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction. NeurIPS, 2025.


W2: Unresolved citation

We apologize for the unresolved citation and thank the reviewer for pointing it out. The revised manuscript now correctly cites Chatzikokolakis et al. (2013), “Broadening the Scope of Differential Privacy Using Metrics,” and includes the corresponding bibliography entry.


Q1 and Q2: Token embeddings and their connection to linear queries

In our experiments, we obtain each token embedding by direct lookup in the pretrained Qwen2.5-7B input embedding table, which remains fixed during training and inference. With E∈R∣V∣×dE\in\mathbb{R}^{|\mathcal V|\times d} and a column one-hot vector xtx_t, the lookup is et=E⊤xte_t=E^\top x_t. The distance between two input tokens t,t′t,t' in our metric-DP definition is the L2 distance between their embeddings: d(t,t′)=∥et−et′∥2.d(t,t')=\|e_{t}-e_{t'}\|_2. We perturb these embeddings and map them back to discrete tokens for the external LLM.

The embedding lookup is a linear query, and existing DP mechanisms for linear queries are relevant to this privatization step. Compatible mechanisms can be incorporated as alternative perturbation mechanisms. However, DP-MIND's effectiveness is not tied to a particular perturbation mechanism, and our contribution lies in DP-aware memory training and local answer denoising for effective long-context inference. Table 2 shows that DP-MIND outperforms direct Qwen-1M inference across dχd_\chi-privacy, SanText, and RANTEXT, with gains of 22.66 to 68.75 percentage points. These results show that DP-MIND improves inference utility across different perturbation mechanisms.

Thank you & Looking forward to your feedback

We would like to sincerely thank you for the time and effort you have devoted to reviewing our work. Your comments have led to concrete additions and clarifications in the revised manuscript. For your convenience, we briefly summarize how we addressed them:

  • W1, comparison with private LLM inference methods. We added comparisons with direct inference and InferDPT using Qwen2.5-7B under the same privacy budget and DP-perturbed inputs. DP-MIND improves Sub-EM by 16.41–25.78 percentage points over InferDPT at context lengths from 7K to 28K. At 56K, where direct inference and InferDPT exceed the model's supported context length, DP-MIND achieves 73.44 Sub-EM through sequential memory processing.

  • W1, comparison with non-private long-context methods. We clarified that the original manuscript already compared against sliding-window attention, recurrent chunk-wise inference, and MemAgent. Following your suggestion, we additionally evaluated FlexPrefill, XAttention, and KVzip under identical DP-perturbed inputs. DP-MIND achieves 68.75–73.44 Sub-EM from 28K to 224K, while these baselines degrade sharply as the context length increases.

  • W2, unresolved citation. We corrected the malformed citation and added the complete reference to Chatzikokolakis et al. (2013), Broadening the Scope of Differential Privacy Using Metrics.

  • Q1, token embedding construction. We clarified that each token embedding is obtained by direct lookup in the fixed pretrained Qwen2.5-7B input embedding table. Equivalently, this operation can be written as multiplying the embedding matrix by the token's one-hot vector.

  • Q2, connection to private linear queries. We agree that embedding lookup is a linear query and that established DP mechanisms for linear queries are applicable. We clarified that DP-MIND is compatible with alternative perturbation mechanisms satisfying the required privacy guarantee. Its main contribution is the DP-aware memory training and local answer denoising that enable effective inference over long perturbed contexts. Our results across three perturbation mechanisms further show that the gains are not tied to one specific embedding perturbation method.

We hope these additions and clarifications address your concerns. If any issue remains unresolved, we would be grateful for further guidance and would be happy to elaborate. Thank you again for your careful and constructive review.

AuthorsOpenReview ↗

A final check before the discussion period closes

Dear Reviewer 4s3j,

As the discussion period approaches its close, we wanted to follow up once more. In response to your comments, we added comparisons with both private inference methods and non-private long-context methods. We also corrected the unresolved citation, clarified that token embeddings are obtained through direct lookup from the fixed Qwen2.5-7B embedding table, and explained the connection between this operation and private linear queries.

We would be grateful to know whether these additions and clarifications address your concerns, or whether any issue remains that we could clarify. We will respond promptly to any further questions during the remaining discussion period.

Thank you again for your time and careful review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.