acceptodds
Under review as a conference paper at ICLR 2027

Soft Prompt Projection to Personalize Across Model Scales

Abstract

As large language models (LLMs) are deployed to real world use cases, it is important that LLMs accurately adapt to individual end users to better act as personal assistants or writing aids. While small LLMs can fit on user devices and train on user data without compromising privacy, devices may need to rely on large server-side LLMs to solve complex tasks, and such models cannot be trained on user data without compromising user privacy. In this work, we present an approach to personalize large server-side LLMs without needing to expose user data to the large model. We first learn personalization parameters for a small LLM on-device and then learn a projection network that maps these parameters to a large model. We show that our method enables large LLMs to personalize to entirely unseen users across tasks including text generation and rating prediction, achieving over 60% F1 on a 5-class rating prediction task where prompting baselines fail to even follow instructions. Our method combines the personalization of a small on-device LLM with the generation capabilities of a large server-side LLM, leading to superior win-rates over a generic large LLM and a personalized small LLM.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
personalization, large language models, soft prompts

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_Ckd5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates cross-model personalization, aiming to transfer user preferences learned by a small on-device language model to a larger server-side model without directly sharing the user's private training data.

Critical strengths

S1. The topic is interesting.

S2. The writing is easy to understand and follow.

S3. Experiments have some interesting insights.

Critical weaknesses

W1. The paper requires a group of training users, and the privacy of these users is not considered.

W2. Another potential limitation is that, it is unknow for the requirement of personalization overlap between training and unseen users. If the distribution various a lot, can the method still works?

W3. For the technical design, using a projection network with a soft prompt for personalization has been studied for a long time. Although the paper may have some new design, the overall framework seems “familiar with not too much surprise” to me.

W4. For the experiments, it is questionable whether, except for the writing style task, other tasks can indicate/need personalization.

Questions for authors

Please address the concerns in weaknesses.

2 comments

We thank the reviewer for their time and consideration.

W1. The paper requires a group of training users, and the privacy of these users is not considered.

We agree, and this is an area we highlight at the end of our conclusion for future work. While private federated learning could be applied to these users, that is beyond the scope of this work. In this work, we show that this architecture works in “ideal conditions”, and future work may consider how to further improve under additional constraints, such as private learning of the projector or global prompts. We will clarify that the training users do share their personal data with the server.

W2. Another potential limitation is that, it is unknow for the requirement of personalization overlap between training and unseen users. If the distribution various a lot, can the method still works?

We test on unseen users in each dataset, ranging from ~300 in MovieLens to ~6700 in AmazonReviews. These unseen users vary significantly, and show that the projector does generalize beyond simply covering the training users. We examine this in more detail by analyzing win rates on the Amazon review writing task for users with varying degrees of similarity to the training users. In the below table, we show win-rates for the Projected method against the GT-Large baseline for users that are within 1 standard deviation of the training-user distribution. 1-2 standard deviations, 2-3 deviations, and 3+ deviations (where distance is computed on the learned small-model soft-prompts).

Model pair0–1 STD+1 STD+2 STD>3 STD
Pythia-410M → Pythia-12b.297.287.265.143
Pythia-410M → Qwen3-8B.364.304.240.252
Pythia-410M → Qwen3-14B.397.309.241.263
Qwen3-0.6B → Pythia-12b.328.292.282.289
Qwen3-0.6B → Qwen3-8B.314.254.243.230
Qwen3-0.6B → Qwen3-14B.320.242.228.213

And comparing the Projected method against the No-Personalization baseline

Model pair0–1 STD+1 STD+2 STD>3 STD
Pythia-410M → Pythia-12b.991.989.988.990
Pythia-410M → Qwen3-8B.986.990.990.986
Pythia-410M → Qwen3-14B.989.992.992.991
Qwen3-0.6B → Pythia-12b.994.988.978.967
Qwen3-0.6B → Qwen3-8B.991.979.966.955
Qwen3-0.6B → Qwen3-14B.994.984.974.961

These results show that performance degrades as prompts are further from the training mean when compared to the GT-Large baseline, but that performance is steady when compared to the No-Personalization baseline.

W3. For the technical design, using a projection network with a soft prompt for personalization has been studied for a long time. Although the paper may have some new design, the overall framework seems “familiar with not too much surprise” to me.

We thank the reviewer for their comment, and we note that we are the first to show this approach generalizes across model scales and families to unseen users, as demonstrated in the “write in my voice” and product recommendation tasks, and we have rigorously evaluated multiple solutions to this problem (e.g., direct mapping, backprop through a large model, global prompts vs. personal prompts, and comparison to prompting baselines). However, we will modify the motivation and conclusion to clarify that this work builds on foundations established in prior work.

Due to character constraints, we continue our response in a separate comment.

W4. For the experiments, it is questionable whether, except for the writing style task, other tasks can indicate/need personalization.

While the Amazon rating prediction task shows little need for personalization, as the task is to predict a rating from review text, such a task may still benefit from personalization. The MovieLens task requires much more personalized learning, as the task is to predict a rating given only a movie title (which means there can be different labels for the same input). Results are given below, showing that F1 performance drops substantially when personal information is removed:

AccuracyMacro-F1
SmallLargeFullGlobal onlyFullGlobal only
Pythia-410MQwen3-14B0.39810.36180.27930.2121
Pythia-410MQwen3-8B0.41140.06570.35400.0247
Pythia-410MPythia-12b0.38230.34600.26170.1574
GT LargePythia-12b0.40650.35650.33870.2050
GT LargeQwen3-8B0.41380.36190.36310.2338
GT LargeQwen3-14B0.42400.36630.37890.2544

In the Projected Qwen3-8B setting, Global only degenerated and only predicted “1” for all samples. These results are over our test set, not the full MovieLens-1M dataset. Note that for the GT Large Global Only setting, there is no pre-pending of zeroed out prompts, and these global prompts were learned across the full training population, meaning they represent the best that this method can possibly do without personalization, and we see that they fall short of both GT Large with personalization and Projected with personalization. These results, coupled with the results in the full paper, show that this personalized extension is beneficial when the task demands personalization, and does not adversely affect task performance when personalization is not required.

Reviewer_CfejOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to personalize a large server-side language model using user data kept on a local device. It trains user-specific soft prompts on a small local model, then learns a projection network that maps those prompts into the large model’s embedding space using data from a separate set of training users. For unseen users, the projected prompts condition the large model without sending their historical data in plain text. The method is evaluated on personalized text generation and rating prediction using Amazon Reviews, Stack Exchange, and MovieLens-1M.

Critical strengths
  • Simple and intuitive method. Learning a projection from a small model’s user-specific soft prompts to a large model’s prompt space is a straightforward way to transfer personalization across model scales. The approach is easy to understand and implement.

  • Extensive empirical analysis. The paper evaluates multiple source and target models, includes several baselines and ablations, and discusses prompt memorization and privacy leakage. These analyses help clarify both the method’s performance and its limitations.

Critical weaknesses

Major issues

  1. The privacy claim is insufficiently validated. The proposed deployment sends the user’s original small-model soft prompt to the server, which could also run the small model or train an attack using data from the projector-training users. The privacy analysis mainly asks the small and large models five questions about the prompts and checks for exact ten-word overlaps with user reviews. This does not assess whether an informed server can infer sensitive attributes or reconstruct user information. Stronger attacks on the actual transmitted prompt are needed to support the privacy claim.

  2. The evaluation covers a narrow range of personalization needs. Although the paper reports extensive results across model pairs and settings, its three datasets support only review or post generation and rating prediction. The authors’ own global-prompt ablation suggests that rating prediction on these datasets requires relatively little user-specific information. This makes it difficult to assess whether the method helps when personalization is genuinely necessary. Evaluation on more diverse generation tasks, such as those in LaMP and LongLaMP, would strengthen the paper.

Minor issues

  1. The related-work discussion is incomplete. The paper discusses soft-prompt transfer but gives limited attention to other privacy-conscious device–cloud personalization approaches, such as CoSteer, P³, and LSRP. Comparing what information these methods transmit to the server would clarify the proposed method’s contribution and privacy assumptions.

  2. The methodological novelty is modest. Soft-prompt tuning and conditioning frozen models with soft prompts are established techniques. The main methodological addition is a learned projection that maps user prompts from a small model to a large one. The paper should position this as an application and extension of existing prompt-transfer techniques, rather than implying a broader methodological advance.

Questions for authors

please see weakness above, these are all critical questions, please answer them accordingly or improve the paper in the future.

5 comments

We thank the reviewer for their thorough review.

Privacy claim is not validated

The primary aim of our work is to ensure that users remain in control of their personal data by keeping in on their device, and that such data are not transmitted to the server. Our paper is primarily a personalization paper, not a formal privacy paper, and we therefore do not provide formal guarantees. However, we are willing to adjust our introduction and conclusion such that we do not over-claim privacy guarantees, but rather that we clearly state our objective of protecting a user’s raw data by keeping it on device. Specifically, we will change the wording of the Figure 1, the abstract, introduction, Sections 4.3 and 5, and the conclusion, such that the emphasis is on the raw text being protected, not the prompt’s information.

Generalization to more complex tasks

One of the central contributions of our work is that we enable personalization without exposing raw data, and that this personalization is possible with unseen users. We therefore consider “write in my voice” tasks (StackExchange and AmazonReviews) and product recommendation (MovieLens and AmazonReviews), which we can construct as personalization tasks with held-out users. Note that these “write in my voice” tasks are reflective of real LLM use-cases, and are personalized tasks in which unique authors have a unique style even on shared products or subjects. Regarding LaMP and LongLaMP, the ICLR guidelines explicitly state that "Data should be used in ways consistent with their licences.", but LongLaMP was not released with a license and LaMP refers all license questions to the original datasets, many of which lack licenses or require commercial agreements. We reached out to the authors while conducting this research and asked for a license to be attached to the LongLaMP dataset so the terms of use are clear. We did not receive a response from the authors, and the dataset remains unlicensed. We therefore consider use of this dataset to be in violation of the ICLR guidelines.

Confounded task and personalization benefits

While we agree that the Amazon rating prediction task shows little need for personalization, as the task is to predict a rating from review text, such a task may still benefit from a personalized model. The MovieLens task requires much more personalized learning, as the task is to predict a rating given only a movie title (which means there can be different labels for the same input). Results are given below, showing that F1 performance drops substantially when personal information is removed:

AccuracyMacro-F1
SmallLargeFullGlobal onlyFullGlobal only
Pythia-410MQwen3-14B0.39810.36180.27930.2121
Pythia-410MQwen3-8B0.41140.06570.35400.0247
Pythia-410MPythia-12b0.38230.34600.26170.1574
GT LargePythia-12b0.40650.35650.33870.2050
GT LargeQwen3-8B0.41380.36190.36310.2338
GT LargeQwen3-14B0.42400.36630.37890.2544

In the Projected Qwen3-8B setting, Global only degenerated and only predicted “1” for all samples. These results are over our test set, not the full MovieLens-1M dataset. Note that for the GT Large Global Only setting, there is no pre-pending of zeroed out prompts, and these global prompts were learned across the full training population, meaning they represent the best that this method can possibly do without personalization, and we see that they fall short of both GT Large with personalization and Projected with personalization. These results, coupled with the results in the full paper, show that this personalized extension is beneficial when the task demands personalization, and does not adversely affect task performance when personalization is not required.

Finally, we note that valid science does not require significant success on all fronts, and we emphasize that our method has enabled personalization across different models scales and families without transmitting raw data, while still outperforming a pure prompting baseline. This is a valid and worthwhile contribution to the community.

Positioning

We thank the reviewer for the suggestion and we will rephrase our motivation and conclusion to clarify that we are not introducing the concepts of soft prompts, and that these are built on foundations established in prior work.

We continue on other address prior work comments in another response due to character limits.

Related work suggestions

We thank the reviewer and will include these in our finished paper. In contrast to our work, CoSteer requires that both models share a tokenizer, and assumes that the “No-Personalization” draft is compatible with the personalized draft from a small model. In cases where there is a significant capabilities gap or personalization gap, this may not be practical. P^3 and LSRP both rely on the small model to be sufficiently strong as a personalized writer or grader, which we show is not the case in our review writing experiments, which motivates our use of a large model as the writer (conditioned on the soft prompt from a small model).

Thanks for your early reply. I have some follow-up questions and i'm willing to raise my score if resolved

As for the privacy claim (W1) and related work inclusion (W3). Your reply is satisfactory, as iclr rebuttal period allows for the direct edit of PDF, please incorporate all this things in your refined version and highlight them in any color you like, W4 is also needed. Thus I would like to see the refined version.

As for the benchmark issue, you point out a copyright issue that i missed before, please re-check the licence becuase there's truly too many works conduct experiments on them, i know a subset of the datasets are not allowed(like lamp-6 and longlamp-1), please check if other subtasks are allowed directly for academic use. If i'm wrong please point out directly and this would not influence your rating.

Thanks for your effort, and i'm looking forward to your future reply.

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

This paper studies personalized transfer across model scales and families, aiming to enable a server-side large model to leverage user preferences already learned by an on-device small model, without acquiring the raw data that the user used for personalized training. The method first learns global soft prompts shared across users, and trains personal soft prompts on the small model using each user's historical data. Subsequently, it uses the data of a set of training users to learn a shared projection network that maps the small model's prompts into the large model's input space. Experiments cover text generation and rating prediction, and demonstrate transferability across model scales and families.

Critical strengths
  1. The method addresses a user privacy issue that is easily overlooked in the current era of large models, and proposes reusing personalized parameters already learned on the device side, so that new users do not need additional optimization on the target large model.
  2. The experiments include multiple model combinations, and split training and test sets by user, which can test generalization to unseen users.
Critical weaknesses
  1. Novelty is limited. CBP-Tuning [1] has already studied shared prompt generators and user-side customization, Plug-in [2] have studied transferring PEFT modules from small models to large models, and PPlug[3] involves projecting user representations into a frozen LLM. This paper has not yet sufficiently clarified its added contribution relative to these lines of work, nor is the comparison with POST sufficient to establish a complete performance advantage.
  2. The benefits of personalization and task adaptation are confounded. Global-only already achieves rating prediction results close to the full method, while directly zeroing out personal prompts may introduce distribution shift and impair task execution ability. Existing experiments have not yet sufficiently demonstrated that the benefits depend on correct user information. The generation evaluation also mainly measures style similarity, and the validation of task quality is relatively limited.
  3. The experiments do not fully validate the stated motivation. This paper motivates the work with personal assistants, writing assistance, and complex requests, but rating prediction has limited representativeness for these scenarios, and only two tasks are mainly covered.

[1] CBP-Tuning: Efficient Local Customization for Black-box Large Language Models, EMNLP2025

[2] Parameter-efficient Tuning for Large Language Model without Calculating Its Gradients, EMNLP2023

[3] LLMs + Persona-Plug = Personalized LLMs, ACL2025

Questions for authors
  1. Compared with existing parameter transfer or user representation projection methods, how large are the advantages of this paper?
  2. After replacing personal prompts with other users’ learned personal soft prompts, how much does performance drop? Compared with a shared prompt independently optimized using the same training data, is the advantage still stable?
6 comments

We thank the reviewer for their thoughtful review.

Confounded task and personalization benefits

We agree that the Amazon rating prediction task shows little need for personalization, as the task is to predict a rating from review text. In contrast, the MovieLens task requires much more personalized learning, as the task is to predict a rating given only a movie title (which means there can be different labels for the same input). Results are given below, showing that F1 performance drops substantially when personal information is removed:

AccuracyMacro-F1
SmallLargeFullGlobal onlyFullGlobal only
Pythia-410MQwen3-14B0.39810.36180.27930.2121
Pythia-410MQwen3-8B0.41140.06570.35400.0247
Pythia-410MPythia-12b0.38230.34600.26170.1574
GT LargePythia-12b0.40650.35650.33870.2050
GT LargeQwen3-8B0.41380.36190.36310.2338
GT LargeQwen3-14B0.42400.36630.37890.2544

In the Projected Qwen3-8B setting, Global only degenerated and only predicted “1” for all samples. These results are over our test set, not the full MovieLens-1M dataset. Note that for the GT Large Global Only setting, there is no pre-pending of zeroed out prompts, and these global prompts were learned across the full training population, meaning they represent the best that this method can possibly do without personalization, and we see that they fall short of both GT Large with personalization and Projected with personalization. These results, coupled with the results in the full paper, show that this personalized extension is beneficial when the task demands personalization, and does not adversely affect task performance when personalization is not required.

Motivation mismatch

We motivate our work with a vision for the future of personal assistants and personalized writing, and we show results on personalized writing and personalized recommendations. To the best of our knowledge, the community currently lacks a personalized agentic benchmark, and the creation of such a benchmark is beyond the scope of this work.

Compared to existing parameter transfer methods

The closest related work to ours is the POST paper from ICML 2025. Compared to this work, our approach is computationally much cheaper and performs better on held-out users (as described in our paper, the POST baseline would have taken over 1600 hours on a 8x80GB A100 node).

After replacing personal prompts with other users’ learned personal soft prompts, how much does performance drop? Compared with a shared prompt independently optimized using the same training data, is the advantage still stable?

We answer the shared prompt question above with the MovieLens-1M experiment. To answer the swapped-prompt question, we generated responses after exchanging user’s soft prompts, but still giving them the same input context. We computed win rates for this Own vs. Swapped setting, as well as win rates against a global-only baseline, which we include below:

Model pairOwn > SwappedOwn > Global onlySwapped > Global only
Pythia-410M → Pythia-12b.502.770.768
Pythia-410M → Qwen3-8B.607.742.655
Pythia-410M → Qwen3-14B.594.684.599

These results show that exchanging prompts in the Pythia→Qwen mapping leads to a significant degradation in performance (nearly a 10% drop in win-rate relative to a global-only prompt) , but that prompts in the Pythia→Pythia mapping are less unique to individual users (nearly 50-50 win rates). This may owe to the difference in the number of personal prompt tokens (given in the Appendix), which is only 1 for the Pythia→Pythia mapping, but is 100 for the Pythia→Qwen mapping.

We continue on to address the prior work comments in a separate response, due to character constraints.

Comparison to prior work

We appreciate the links to prior work and will include them in our final paper. While relevant, these papers do not directly address all of our core contributions.

  • CBP [1] requires sending a user’s personal data to the large model for the prompt optimization procedure (for example, sending context, prior examples, or other raw text). As stated in our work, the objective of our method is to avoid the need to send any unnecessary raw text to the server.
  • Plug-in [2] requires training a “Bridge” for every new task, and this is a form of distilled larger model. This setup, akin to the POST baseline that we consider, is significantly more expensive than our approach, and does not permit arbitrary small/large model pairs (the “Bridge” must be in the same model family as the large model, and must share a tokenizer). It is also not clear that this method would generalize to a personalization setting, with diverse users in the same task.
  • PPlug [3] composes a custom personal prompt for each new query, as opposed to learning a single generalized personalization prompt as in our work, and this prompt is aggregated over prior data and context, which we do not assume access to at test-time. Additionally, PPlug does not consider generalization to entirely unseen users. We will include these works in our related works section with descriptions of how they relate to our paper.

Thank you for the additional experiments and clarifications. However, your response does not fully alleviate my concerns.

For W1, the distinctions from those works clarify the different assumptions, but they do not fully resolve the conceptual overlap. I do not see a clear substantive distinction at the level of the core problem or intended application. These methods all aim to use locally derived user or task information to personalize a server-side or frozen LLM under privacy and efficiency constraints. The differences appear mainly to concern implementation details and experimental assumptions. Please clarify what core technical contribution is missing from prior work, rather than only describing differences in pipeline or evaluation protocols.

For W2, the new MovieLens global-only baseline and own-versus-swapped results partially address this concern. The cross-family results are encouraging, although the near-chance Pythia-to-Pythia result and the explanation based on prompt length remain less conclusive.

For W3, I agree that introducing a new agentic benchmark is beyond the scope of this work. However, the current experiments still cover a relatively narrow set of tasks, so the evidence for broader personal-assistant or complex-query scenarios remains limited.

Comments on the paper

1 comment

We appreciate the reviewers’ engagement and discussion on our work.

Following reviewer suggestions, we have clarified our problem statement and paper scope, which now more accurately reflects both the privacy claims and the datasets in the paper. We are happy to make further modifications, as mentioned in our discussion with reviewer FpMF, if these are deemed useful by the reviewers. Along these lines, we have also added related works that were suggested by the reviewers, along with clearer positioning with respect to these works. Finally, we have also added experiments that provide more rigorous empirical evidence for our claims around the utility of persona vs. global prompt parameters.

These changes have helped to strengthen our work, and we are grateful to the reviewers for their time and support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.