acceptodds
Under review as a conference paper at ICLR 2027

Personalization by Completion: Collaborative Filtering in LLM Weight Space

Abstract

Large language models increasingly interact with users who have different preferences, goals, and styles. Personalizing these models is difficult because a user's interaction history is often sparse: many preferences that matter for a new query may never have been observed. We view this problem as a form of collaborative filtering. Just as recommendation systems infer missing entries in a user–item matrix from patterns across users, we infer how a language model should adapt to a new user and topic from adaptation patterns learned across a population. Our key idea is to represent these patterns directly in the model's low-rank adaptation weights and compose a conditional weight update for an unobserved user–topic pair. The resulting model can personalize its behavior even when the relevant preference is absent from the user's history. Experiments across preference prediction tasks and language models show that this approach consistently improves personalization, with gains that persist even when the full observed history is available. These results suggest that personalized language models can learn not only from what a user has said, but also from what similar users reveal.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Personalization, LLM, Recommendation Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_t34YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of completing missing preferences in the context of large model personalization and proposes CP-gated LoRA. Combining collaborative filtering with LoRA, the method first extracts feature representations for each user from the records of the entire user base; it then dynamically adjusts shared LoRA weights based on the current topic and model layer, thereby leveraging information from similar users to make predictions. In terms of experiments, the authors constructed a synthetic dataset named PersonaGrid and conducted tests using multiple open-source models and the MovieLens dataset; the overall performance surpassed baselines that incorporate personal history or summaries into the prompt. Additionally, the effectiveness of the method was validated through experiments involving variations in observation density and the substitution of user representations.

Critical strengths
  1. The issue raised by the authors is indeed significant and has often been overlooked; the introduction is well-written and clearly articulates the importance of studying this problem.

  2. The problem statement is very clear, explicitly outlining the task setup and evaluation objectives.

  3. In addition to the main experiments, the authors designed intervention experiments--such as substituting user representations--that directly validate the actual role user representations play in the prediction process.

Critical weaknesses
  1. The baseline comparisons are not entirely fair. While ICL and Memory-ICL rely solely on the current user's own history, the authors' method utilizes data from other users. Consequently, it is unclear whether the performance gains stem from the modification of LoRA weights or simply from the inclusion of external information. The experiments should incorporate baselines that also leverage data from other users—such as collaborative filtering, prompt-based inclusion of similar users' information, or comparisons with methods like CoRA.

  2. The experimental setup is overly idealized. The synthetic datasets were generated based on a few fixed prototypes, naturally aligning with the paper's approach of clustering users beforehand. Furthermore, the experiments were limited to closed-ended multiple-choice questions, omitting open-ended text generation tasks. This makes it difficult to determine whether the method remains effective in real-world scenarios involving personalized generation with large models.

Questions for authors
  1. How does the proposed method--which modifies LoRA weights--compare to an approach where information from other users (such as records of similar users or collaborative filtering results) is directly incorporated into the prompt? What are the specific advantages of embedding this information into the weights versus simply including it in the prompt?

  2. For the tasks evaluated in the paper, what kind of performance could be achieved by using traditional collaborative filtering algorithms directly, without a large language model? Under what circumstances is it necessary to encode collaborative filtering directly into the LoRA weights, rather than using collaborative filtering to calculate missing preferences first and then feeding that information to the large language model?

  3. The current experiments are structured as multiple-choice tasks. Would this method remain effective if the model were required to generate free-form, full-paragraph responses or engage in everyday conversation? Could modifying the LoRA weights compromise the model's inherent quality of expression?

  4. If a user has previously expressed relevant preferences--or if their tastes differ significantly from the majority of their group--how ​​does the model decide whether to rely on that user's history or the group's aggregate results? How can the model avoid being skewed by the group's mainstream opinion?

4 comments
AuthorsOpenReview ↗

Thank you for your review, we addressed your questions below

We sincerely thank Reviewer t34Y for recognizing the importance of the problem, the clarity of the Introduction and the representation-intervention evidence. We address the questions below:

W1: Source of the performance gains

Sorry for the confusion about the source of the gains. We clarify that collaborative filtering is the key to the gains, instead of the modification of LoRA weights. Following the reviewer's suggestion, we added baselines that also leverage data from other users.

  • Collaborative filtering improves accuracy as text, as soft tokens and as LoRA weights.
MethodPersonaGrid, seenPersonaGrid, unseen
ICL, current user's own history29.3%29.2%
Prompt-based inclusion of similar users' information27.3%25.8%
Collaborative filtering, as text in the prompt54.8% (+25.5%)50.7% (+21.5%)
Collaborative filtering, as soft tokens54.8% (+25.5%)51.4% (+22.2%)
Collaborative filtering, as LoRA weights55.2% (+25.9%)50.7% (+21.5%)
  • Our CP-gated LoRA is more compact than CoRA. Ours uses about 1/120 of CoRA's conditional parameters.
MethodConditional parametersML-1M, seenML-1M, unseen
CoRA40M64.2%64.8%
Our CP-gated LoRA0.33M64.5%63.4%

W2: Effectiveness in real-world scenarios

  • Our CP-gated LoRA improves by +6.2% to +9.9% over ICL on real ML-1M and ML-25M data without prototypes.
DatasetSeenUnseen
ML-1M movies58.2% → 65.3% (+7.1%)59.4% → 69.3% (+9.9%)
ML-25M movies61.2% → 67.4% (+6.2%)60.3% → 66.9% (+6.6%)
  • In open-ended generation, our CP-gated LoRA's responses match the user's preference more often than ICL's. Win/tie/loss is 36/49/15 on PersonaGrid.

Q1: Advantages of weights > prompt

  • As clarified in W1, collaborative filtering is the key, and LoRA weights are one way to use it. Including the collaborative filtering results in the prompt reaches accuracy similar to embedding them into the weights, while including records of similar users in the prompt gives no gain (table in W1).

  • Embedding the information into the weights takes fewer tokens and needs less downstream data.

    • It takes only about 1/20 of the prompt tokens of including similar users' records in the prompt.
    • With 10% of the downstream data, accuracy improves by +7.7% for seen users and +6.4% for unseen users over soft tokens in the prompt.
MethodPrompt tokens, seenPrompt tokens, unseen
Similar users' records in the prompt259340
Our CP-gated LoRA15 (-94.2%)15 (-95.6%)
MethodTraining dataML-1M, seenML-1M, unseen
Soft tokens in the prompt10%53.1%53.8%
Our CP-gated LoRA10%60.8% (+7.7%)60.2% (+6.4%)

Q2: Necessity of the LLM and weights beyond traditional collaborative filtering

  • Our CP-gated LoRA brings the accuracy of traditional CF into the LLM's answers. Without an LLM, MF and user-kNN output a rating, so they cannot write the free-form responses evaluated in Q3. On the same test cells, CP-gated LoRA is within 0.6 points of MF, above user-kNN, and 6.2–9.9 points above ICL.
MethodML-1M movies, seenML-1M movies, unseenML-25M movies, seenML-25M movies, unseen
ICL58.2%59.4%61.2%60.3%
User-kNN63.4%63.5%65.6%65.4%
Direct MF65.8%69.5%67.9%67.5%
Our CP-gated LoRA65.3%69.3%67.4%66.9%
  • Weights help when the completed preference is continuous rather than one group-level answer. On ML-1M genres, our CP-gated LoRA reaches 64.5/63.4, versus 57.5/60.0 when the computed preference is fed as text.

Q3: Effectiveness in free-form generation and quality of expression

  • Yes, our CP-gated LoRA is more effective than ICL for free-form, full-paragraph responses. Its responses match the user's preference more often. A separate, larger LLM judge compares the two responses against the user's true preference.
UsersWinTieLoss
PersonaGrid, seen36%49%15%
PersonaGrid, unseen33%48%19%
MovieLens, unseen24%57%19%
  • Adding the collaborative information to the LoRA weights keeps the quality of expression. On a 100-instruction probe, responses with it win 24%, tie 57% and lose 19% against responses without it.

Q4: Conflict between the user's history and the group's mainstream opinion

  • The model does not have to choose, because a user's group is made of users with similar histories. By construction, collaborative filtering guarantees that the user's history does not differ too much from the majority of its group. The group is used only for preferences that the history does not cover.

  • When the user's history disagrees with the group, the model follows the user. With two relevant ratings in the history, our CP-gated LoRA answers according to the user in 93.9% of such cases for seen users and 91.1% for unseen users, so it is not skewed by the group's mainstream opinion.

Please let us know if you have additional questions.

AuthorsOpenReview ↗

Follow-up on our response and the revised paper

Dear Reviewer t34Y,

Thank you again for the detailed review. Our response above covers the source of the performance gains (W1), real-world effectiveness (W2), and Q1 to Q4. The revised PDF now includes these results.

We would appreciate hearing whether the response resolves these concerns, and we are happy to provide further clarification during the discussion period.

Thanks to the authors for the response and the updated paper. I have read the rebuttal carefully and see the points the authors are making. Even so, my overall view of the paper stays the same, so I will keep my score.

AuthorsOpenReview ↗

Could the reviewer point to the remaining concerns?

Thank you for reading the response and the revised paper.

To make the remaining discussion period useful, could the reviewer point to the specific concerns that are still unresolved? The response addressed the source of the performance gains (W1), real-world effectiveness (W2), and Q1 to Q4, each with new experiments now included in the revision. If any of these answers is unconvincing, or if a concern lies outside these items, we would be glad to address it directly.

Reviewer_BWDEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on whether LLMs can infer user preferences that did not emerge in historical context with collaborative filtering signals. They first obtain user representations with CF, and then employ CP-gated LoRA to combine user, effective topic, and adapter-target factors to jointly control model weight updates. The experiments demonstrate that this approach outperforms ICL that relies solely on user history.

Critical strengths
  1. The motivation to introduce CF signals into LLM inference is reasonable and important. Inspired by applications in recommender systems, the paper achieved significant performance improvements in specific LLM scenarios.

  2. The experiments cover multiple LLM backbones, and extensive experiments demonstrate the effectiveness of the proposed method compared to ICL.

Critical weaknesses
  1. The experiments only compares ICL and the proposed method, without addressing potentially closer methods such as CoRA and CoLLM. Introducing CF signals into LLM inference is not a new topic. Although the authors cite these works and provide brief comparative analyses, a more appropriate comparison might be directly realizing the core designs of these related works on the current task and reporting the performances.

  2. The theoretical guidance of the CP-gated LoRA design is relatively weak. The paper provides an intuitive explanation of the three-factor decomposition, but lacks further theoretical analysis explaining why it is more suitable for completing preferences. The reviewer concerns that the current understanding of this design might only be heuristic.

  3. The experimental results mainly focus on candidate answer prediction, which may not fully demonstrate improvements in generation quality or user satisfaction. Furthermore, the paper does not provide latency results or the cost of constructing upstream representations.

Questions for authors

Please refer to Weakness.

2 comments
AuthorsOpenReview ↗

Thank you for your review, we addressed your questions below

We sincerely thank Reviewer BWDE for recognizing the importance of collaborative filtering for personalization and the results across multiple backbones. We address the questions below:

W1: Ablation study on comparison with CoRA and CoLLM

Following the reviewer's suggestion, we realized the core designs of CoRA and CoLLM on our task. With the full data, their accuracy is similar to our CP-gated LoRA.

  • Our CP-gated LoRA is more compact and needs less downstream data. It uses about 1/120 of CoRA's and 1/8 of CoLLM's conditional parameters. With 10% of the data, accuracy improves by +7.7% for seen users and +6.4% for unseen users over CoLLM.
MethodConditional parametersTraining dataML-1M, seenML-1M, unseen
CoRA40M100%64.2%64.8%
Our CP-gated LoRA0.33M100%64.5%63.4%
MethodConditional parametersTraining dataML-1M, seenML-1M, unseen
CoLLM2.64M10%53.1%53.8%
Our CP-gated LoRA0.33M10%60.8% (+7.7%)60.2% (+6.4%)
  • Our way of using the collaborative information is different from CoRA and CoLLM. We multiply a user factor, a topic factor and an adapter-target factor to weight shared LoRA directions. CoRA trains a 40M-parameter network to generate weight updates, and CoLLM maps the collaborative information to soft tokens in the input.

W2: Theoretical guidance of the three-factor design

As mentioned in the W1 response, our CP-gated LoRA reaches similar accuracy with a smaller model and less data. The three-factor decomposition gives the design a structure. It uses the same products of user and topic factors as matrix factorization, so the same factors serve user–topic combinations absent from training. Our paper is empirical, and theoretical analysis is out of the scope of this paper.

W3: Generation quality, latency and cost of upstream representations

  • In free-form generation, our CP-gated LoRA's responses match the user's preference more often than ICL's. Win/tie/loss is 36/49/15 and 33/48/19 for seen/unseen PersonaGrid users.

  • Our CP-gated LoRA is faster than ICL. With a batch size of 32, ours takes 29.1 ms per query, versus 105.5 ms for ICL.

  • Constructing the upstream representations takes seconds. Offline on CPU, ours takes 3.3 s for PersonaGrid and 8.6 s for ML-1M.

Please let us know if you have additional questions.

AuthorsOpenReview ↗

Follow-up on our response and the revised paper

Dear Reviewer BWDE,

Thank you again for the review. Our response above covers the comparison with CoRA and CoLLM (W1), the guidance behind the three-factor design (W2), and generation quality, latency and cost (W3). The revised PDF now includes these results.

We would appreciate hearing whether the response resolves these concerns, and we are happy to provide further clarification during the discussion period.

Reviewer_WRSDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of sparse preference completion in LLM personalization. When one target user’s preference information is missing in his/her history record, how does the LLM use other users’ preference information to infer the missing preference and generate personalized response. The authors propose CP-gated LoRA, first constructs the user representation by using collaborative filtering, then dynamically updates the LoRA weights by using the user \textit{u}’s representation, the effective topic factor, and the adapter-target factor. The authors conduct the experiment on synthetic dataset PersonaGrid and real-world dataset MovieLens. Results show this method performs better than user-history based methods like ICL and Memory-ICL on sparse preference completion task and could be generalized to users that have not been trained by the LLM.

Critical strengths

S1. The research question has practical significance. This paper focuses on completing the missing preference information in user history, rather than merely improving the utilization of known historical information, which has practical value for LLM personalization in sparse feedback scenarios. S2. The method design is reasonable. The authors combine group collaborative information with parameter-efficient fine-tuning, controlling low-rank weight updates through users, topics, and Adapter positions jointly, avoiding the need to maintain separate Adapters for each user or user-topic combination. S3. The experiments provided valuable observations. The authors proposed the PersonaGrid Benchmark, which specifically test the condition in S1, other than that, authors also used real world dataset, MovieLens to validate the method across multiple backbone models and different user divisions.

Critical weaknesses

W1. Missing critical baselines to disentangle the contribution of population information from that of the proposed CP-gated LoRA architecture.

The main baselines, ICL and Memory-ICL, uses target user’s own history and a shared LoRA adapter without the user representation inferred by the collaborative filtering, making it unclear whether the improvements come from the additional collaborative information or the proposed weight adaptation mechanism. Several important controls are missing: (1) cluster-majority, matrix factorization, or MLP predictors using the same user representations in CP-gated LoRA (2) a population-average baseline and a two-stage preference completion followed by textual prompting, as explored in PEP/CAPE [1] (3) collaborative filtering and input-space conditioning baselines considered in CoRA [2] and (4) cross-user retrieval-based ICL explored in PersonalLLM [3]. Adapted comparisons with matched information access would help establish the necessity of CP-gated LoRA. References:

[1] Bose, A., et al. Cold-Start Personalization via Training-Free Priors from Structured World Models. Lifelong Agent Workshop at ICLR, 2026. arXiv:2602.15012.

[2] Liu, Y., et al. CoRA: Collaborative Information Perception by Large Language Model's Weights for Recommendation. AAAI, 2025.

[3] Zollo, T. P., et al. PersonalLLM: Tailoring LLMs to Individual Preferences. ICLR, 2025.

W2. The underlying methodological ideas to solve the research problem substantially overlap with prior work, while the distinctions from related studies are insufficiently established.

Inferring unobserved preference dimensions from group correlations, including cases where extensive user history does not cover the relevant preferences, has already been investigated by PEP/CAPE [1]. CoRA [2] generates low-rank weight updates from collaborative filtering representations and integrates them directly into a frozen LLM. Moreover, PersonalLLM [3] explicitly studies cross-user learning when user history is limited or lack of information relavent to the current query, including cold-start scenarios involving new topics (Section 4.2). While the present work introduces a specific user–topic–adapter factorization and explicitly evaluates missing preference dimensions, the paper does not convincingly isolate the technical advantage of this parameterization over existing collaborative weight adaptation mechanisms.

References:

[1] Bose, A., et al. Cold-Start Personalization via Training-Free Priors from Structured World Models. Lifelong Agent Workshop at ICLR, 2026. arXiv:2602.15012.

[2] Liu, Y., et al. CoRA: Collaborative Information Perception by Large Language Model's Weights for Recommendation. AAAI, 2025.

[3] Zollo, T. P., et al. PersonalLLM: Tailoring LLMs to Individual Preferences. ICLR, 2025.

W3. The design of the PersonaGrid benchmark is closely aligned with the method's representation, while the evaluation remains restricted to structured preference classification.

PersonaGrid uses eight population prototypes, and the proposed method represents users using eight active cluster indicator coordinates, with observed answers parsed according to the benchmark's predefined factor schema. This close alignment between the benchmark construction and the model representation raises concerns about whether the reported improvements generalize to less structured preference distributions or alternative representation choices. Both PersonaGrid and the MovieLens genre preference prediction [4] ultimately measure candidate classification rather than open-ended personalized generation. In contrast, PEP/CAPE [1] evaluates the quality of generated personalized responses using the PrefDisco benchmark [5], while PersonalLLM [3] evaluates how well generated responses align with individual users' preferences. Other than those methods, the LaMP [6] provides diverse real world personalization tasks, including both personalize classification and personalize text generalization task types. It remains unclear whether the proposed approach improves general LLM personalization other than structured preference prediction without evaluation on more diverse personalization benchmarks.

References:

[1] Bose, A., et al. Cold-Start Personalization via Training-Free Priors from Structured World Models. Lifelong Agent Workshop at ICLR, 2026. arXiv:2602.15012.

[3] Zollo, T. P., et al. PersonalLLM: Tailoring LLMs to Individual Preferences. ICLR, 2025.

[4] Harper, F. M., and Konstan, J. A. The MovieLens Datasets: History and Context. ACM Transactions on Interactive Intelligent Systems, 2015.

[5] Li, S. S., et al. PrefDisco: Benchmarking Proactive Personalized Reasoning. ICLR, 2026.

[6] Salemi, A., Mysore, S., Bendersky, M., and Zamani, H. LaMP: When Large Language Models Meet Personalization. ACL, 2024, pp. 7370–7392.

Questions for authors

Q1 (W1). Could the authors include cluster majority, matrix factorization, and simple MLP predictors using the same user representations inferred from population data? Comparisons with population average predictions, PEP/CAPE style preference completion followed by textual prompting, CoRA style conditional weight generation, and PersonalLLM style cross-user retrieval would also be informative, provided that the methods receive comparable information and supervision.

Q2 (W2). Given that PEP/CAPE already studies the inference of unexpressed preferences and CoRA already injects collaborative filtering information into LLM weights, what is the specific methodological contribution of CP-gated LoRA beyond these approaches? Which experiment directly isolates and demonstrates this contribution? Given that PersonalLLM also investigates cross-user personalization under sparse user histories, could the authors clarify how their problem setting differs from this prior work?

Q3 (W3). How sensitive are the results to the number of inferred clusters, particularly when K differs from 8, or when user representations are constructed without the predefined factor schema? Could the authors evaluate the approach on PrefDisco, PersonalLLM, or LaMP using a carefully constructed missing preference split and assess the quality of open-ended personalized responses rather than only candidate classification accuracy?

2 comments
AuthorsOpenReview ↗

Thank you for your review, we addressed your questions below

We sincerely thank Reviewer WRSD for recognizing the practical importance of missing-preference completion and the shared-adapter design. We address the questions below:

Q1 (W1): Ablation study on collaborative information vs. weight adaptation mechanism

Sorry for the confusion about the source of the improvement. We clarify that the additional collaborative information is the key across all reviewer-suggested forms, instead of the weight adaptation mechanism.

Following the reviewer's suggestion, we added the cluster-majority and MLP predictors and compared each with and without our collaborative information. Both improve by 8.5 to 29.1 points once it is added, so the predictor design is not key.

(1) Accuracy (%) gained from collaborative information.

PredictorPersonaGrid, seenML-1M, seenPersonaGrid, unseenML-1M, unseen
Cluster-majority27.9% → 55.8% (+27.9%)48.8% → 57.3% (+8.5%)23.9% → 51.5% (+27.6%)47.0% → 59.7% (+12.7%)
MLP28.6% → 55.2% (+26.6%)51.5% → 65.5% (+14.0%)22.4% → 51.5% (+29.1%)52.7% → 63.4% (+10.7%)
CP-gated LoRA55.2%64.5%50.7%63.4%

(2) The collaborative information improves accuracy through both text and weights. On PersonaGrid, two-stage completion followed by textual prompting and CP-gated LoRA have similar accuracy, about +27 points above the population average without the collaborative information.

(3) Collaborative information is also the key to improving other forms of adaptation. With the additional collaborative information, both CoRA and CoLLM improve on ML-1M.

MethodUser representationSeenUnseen
CoRA-style conditional weight generationShuffled across users49.6%50.9%
CoRA-style conditional weight generationOur CF representation64.2% (+14.6%)64.8% (+13.9%)
CoLLM-style input-space conditioningShuffled across users49.4%51.0%
CoLLM-style input-space conditioningOur CF representation64.2% (+14.8%)61.8% (+10.8%)
CP-gated LoRAOur CF representation64.5%63.4%

(4) Collaborative information improves accuracy over cross-user retrieval-based ICL on PersonaGrid.

MethodPersonaGrid, seenPersonaGrid, unseen
Cross-user retrieval-based ICL27.3%25.8%
Our CP-gated LoRA55.2% (+27.9%)50.7% (+24.9%)

Q2 (W2): Ablation study on methodological contribution

We appreciate the opportunity to further clarify our methodological contribution. As shown in Q1 (W1), our contribution lies in treating LLM personalization as completing a user–topic preference matrix, instead of in the weight parameterization.

  • Comparison with PEP/CAPE: We complete the preference matrix by collaborative filtering and use the result to gate shared LoRA weights, while PEP/CAPE infers unexpressed preferences and states them as text in the prompt.

  • Comparison with CoRA: We use about 1/120 of CoRA's conditional parameters as a compact gate and reach comparable accuracy, while CoRA generates weight updates with a 40M-parameter network.

MethodConditional parametersML-1M, seenML-1M, unseen
CoRA-style conditional weight generation40M64.2%64.8%
Our CP-gated LoRA0.33M64.5%63.4%
  • Comparison with PersonalLLM: We infer the missing preference by collaborative filtering and use about 95% fewer prompt tokens than PersonalLLM-style retrieval of similar users' records. Our setting also differs: the preference being asked about never appears in the user's own history.
MethodPrompt tokens, seenPrompt tokens, unseenPersonaGrid, seenPersonaGrid, unseen
PersonalLLM-style retrieval of similar users' records25934027.3%25.8%
Our CP-gated LoRA15 (-94.2%)15 (-95.6%)55.2% (+27.9%)50.7% (+24.9%)

Q3 (W3): Sensitivity to KK and the factor schema; open-ended personalized generation

Following the reviewer's suggestion, we added four analyses: the number of inferred clusters KK, user representations without the predefined factor schema, real data without prototypes, and open-ended personalized responses.

  • Accuracy is sensitive only to KK and is highest at K=8K = 8.

  • Accuracy remains similar with or without the predefined factor schema.

KKWith schema, seenWithout schema, seenWithout schema, unseen
434.6%34.1%29.4%
853.2%55.6%50.6%
1638.7%40.0%36.8%
  • Collaborative information improves accuracy on real data without prototypes.
DatasetSeenUnseen
ML-1M movies58.2% → 65.3% (+7.1%)59.4% → 69.3% (+9.9%)
ML-25M movies61.2% → 67.4% (+6.2%)60.3% → 66.9% (+6.6%)
  • In open-ended personalized generation, our CP-gated LoRA's responses match the user's preference more often than ICL's. A separate LLM judge compares the two responses against the user's true preference.
UsersWinTieLoss
PersonaGrid, seen36%49%15%
PersonaGrid, unseen33%48%19%
MovieLens, unseen24%57%19%

Please let us know if you have additional questions.

AuthorsOpenReview ↗

Follow-up on our response and the revised paper

Dear Reviewer WRSD,

Thank you again for the detailed review. Our response above covers the ablation on collaborative information vs. the weight adaptation mechanism (W1), the methodological contribution (W2), and the sensitivity to KK and the factor schema together with open-ended generation (W3). The revised PDF now includes these results.

We would appreciate hearing whether the response resolves these concerns, and we are happy to provide further clarification during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.