PrefReward: Learning User Preference Matrix for Personalized Text Generation

Chengbing Wang; Fuli Feng; Xiaoyan Zhao; Yang Zhang; Yimeng Bai; Yue Wu

REVIEW 5 major objections 6 minor 15 references

PrefReward claims that scoring candidate texts by KL divergence from a user's explicit style-preference matrix personalizes LLM output at decode time, with no fine-tuning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 08:31 UTC pith:POXXJHGI

load-bearing objection A train-free personalization idea with an interpretable preference matrix, but the core token-to-label mapping is left undefined, the closest baseline is missing, and the empirical gains are thin. the 5 major comments →

arxiv 2607.21067 v1 pith:POXXJHGI submitted 2026-07-23 cs.CL

PrefReward: Learning User Preference Matrix for Personalized Text Generation

Yue Wu , Chengbing Wang , Yimeng Bai , Xiaoyan Zhao , Yang Zhang , Fuli Feng This is my paper

classification cs.CL

keywords personalized text generationuser preference matrixKL divergence rewarddecoding-time personalizationpreference interpretabilitytraining-free personalizationstyle alignment

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that personalized text generation can be achieved at decoding time by building an explicit, human-readable preference matrix for each user from token-level probabilities over a fixed set of style labels, then selecting among candidate outputs the one whose style vector is closest to the user's average. This matters because it offers a training-free, interpretable route to personalization, avoiding opaque fine-tuned parameters and retrieval noise. The paper reports consistent gains over non-personalized and retrieval-based baselines on a long-context personalization benchmark using two backbone LLMs of different sizes.

Core claim

On the paper's own terms, the central discovery is that a user's stylistic tendencies can be summarized by averaging, over all tokens in the user's historical texts, the probabilities the backbone LLM assigns to a predefined set of 41 stylistic labels; aggregating these per-text vectors yields a user preference matrix. At generation time, the method samples several candidate outputs and picks the one whose preference vector has minimum KL divergence from the user's mean preference distribution. The paper reports that this explicit-matrix-plus-KL-reward procedure consistently outperforms non-personalized and retrieval-based baselines on all three evaluation metrics, for both a zero-shot backb

What carries the argument

The central object is the user preference matrix P_u, a K-by-N matrix where each row is a per-historical-sample preference vector whose entries are the average probabilities the backbone LLM assigns, over all tokens of that sample, to N = 41 predefined stylistic labels such as formal, concise, vivid, and humor. The reward mechanism is the negative KL divergence between a candidate's preference vector and the user's mean preference vector; a Best-of-N sampling loop with this reward is what steers generation toward the user's style.

Load-bearing premise

The whole framework rests on the assumption that the average of token-level logit probabilities for 41 hand-picked style labels, taken from a general-purpose LLM, faithfully captures a text's style, so that a candidate close in KL divergence to the user's average is genuinely what the user would prefer.

What would settle it

Take a user whose writing style is dominated by a dimension absent from the 41-label set (e.g., heavy emoji use or extremely terse sentence fragments) and run PrefReward with candidates that clearly match versus clearly violate that style; if the KL reward cannot rank the matching candidate above the violating one (or ranks randomly), the style-representation assumption fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Personalization becomes inspectable: the preference matrix can be visualized and compared across users, unlike latent style embeddings.
The method transfers to any pretrained LLM without extra training, since it only uses logits available during forward passes plus a decoding-time selection rule.
Long-context user histories are handled more reliably than retrieval-based methods, because the matrix summarizes persistent style rather than depending on noise-prone retrieved snippets.
Increasing the number of style labels improves quality up to a point; the paper finds roughly 20 labels give a good balance between expressiveness and efficiency.
If correct, retrieval-based personalization can be replaced or supplemented by a cheaper, explicit reward signal for stylistic alignment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same preference matrix could be repurposed beyond candidate selection, for example as a per-user steering vector in activation-space methods or as a prior for fine-tuning, since it already quantifies which of the 41 dimensions a user favors.
The fixed 41-label set is the real bottleneck: users whose style falls outside those labels would get near-constant preference vectors, making the reward blind; an obvious extension is to expand or learn the label set per user.
Because the reward compares only relative style, it could be combined with a content-quality reward such as coherence or task success, to avoid sacrificing factual accuracy for stylistic match.
One could test whether a user's preference matrix transfers across tasks (e.g., review writing vs. email drafting) and still ranks candidates consistently; the paper only evaluates one task.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A train-free personalization idea with an interpretable preference matrix, but the core token-to-label mapping is left undefined, the closest baseline is missing, and the empirical gains are thin.

read the letter

The core idea is worth a conversation: represent a user's style as an explicit matrix of preference-label scores extracted from LLM token logits, then use a KL-divergence reward to pick among best-of-N candidates. That is a clean, interpretable, train-free framing, and the paper gets credit for stating it clearly. The LongLaMP results are plausible and show modest gains over RAG and CoS on both backbones, with a nice case study.

The problems are serious, though. The stress-test is right: Section 3.1 says the preference vector is built by keeping only tokens that belong to P, where P is a fixed set of abstract labels (formal, concise, vivid, humor). No lexicon, embedding projection, or classifier is given. Literally taken, the method would search for the surface tokens 'formal' and 'humor' in the vocabulary, which would be rare and meaningless. This is not a reproducibility nit; it is the load-bearing mechanism. Both the user profile vector and each candidate's vector are built this way, so the KL reward may be computed on an arbitrary projection. The paper is not acceptable until this mapping is defined and validated.

Second, the closest prior work, Drift, is invoked as inspiration but never included as a baseline. That is a glaring omission, particularly because the preference-label set and decoding-time approach are taken from Drift. Without a Drift comparison, the incremental contribution is impossible to assess.

Third, the empirical support is thin. One dataset, no error bars or significance tests, and the zero-shot ROUGE-1 gain over NonPers is 0.002. The aligned gain over RAG-CT is about 0.011, which could be real but needs confidence intervals. Also, Section 5.1.1 says profile samples are randomly drawn with K=2, while Section 3.1 describes BM25-based selection of top-K. That inconsistency matters because K=2 is a very small basis for a preference matrix.

What saves this from rejection-in-a-vacuum is that the idea is sound in principle and could be made solid with explicit mapping, a Drift baseline, and proper statistical reporting. As written, I would not cite or rely on it.

Send it to review, because the framing is useful and a capable reviewer could push the authors toward a reproducible version. But the decision should be a major revision, not acceptance.

Referee Report

5 major / 6 minor

Summary. The paper proposes PrefReward, a training-free decoding-time personalization method. It constructs a user-specific preference matrix by running the backbone LLM token-by-token, filtering the next-token logits to a fixed set of 41 stylistic labels, and averaging the resulting vectors over retrieved profile texts. At generation time, PrefReward samples N_c candidate outputs and selects the one with the highest negative KL divergence between the candidate's preference vector and the user's mean preference vector. Experiments on the LongLaMP personalized review-writing task compare the method against NonPers, BM25 RAG, Contriever RAG, and Context Steering baselines on Llama-2-7B-Chat and an aligned Gemma-2B-IT model. The paper claims consistent improvements in both generation quality and interpretability.

Significance. If the core mechanism were precisely defined, PrefReward would offer a lightweight and interpretable personalization signal that operates entirely at decoding time without fine-tuning. The explicit preference matrix and KL-divergence reward are conceptually appealing and potentially useful for controllable generation. However, the current manuscript leaves the central token-to-label mapping unspecified, which makes the preference vectors and the reward formally undefined. The empirical support is also thin: the zero-shot gains over the non-personalized baseline are around 0.002 ROUGE-1, no error bars or significance tests are reported, only one dataset is used, and the claimed interpretability is not measured. The paper also provides no code or release, which further limits reproducibility. The idea is salvageable, but the present version is not sufficient for publication.

major comments (5)

[Section 3.1, 'Token-Level Preference Extraction'] The definition of v_i is not well-formed. The text says logits are 'filtered to retain only tokens that belong to P', but P is a set of abstract stylistic labels such as formal, concise, vivid, humor. No mapping from vocabulary tokens to these labels is provided: there is no lexicon, no embedding similarity, no classifier, and no manual annotation. If the filter literally matches surface token strings such as 'formal', the resulting vector will be dominated by rare lexical occurrences and will not represent a stylistic distribution. If an implicit semantic mapping is intended, it must be specified. Because both the user mean preference vector \bar P_u and each candidate's p_y(i) are constructed by this operation, the reward R_i = -KL(p_y(i) || \bar P_u) is computed on an arbitrary and unvalidated projection. This is the load-bearing mechanism of the paper, and it must be made concrete an
[Section 5.1.1 vs Section 3.1] The profile-selection procedure is described inconsistently. Section 3.1 defines a BM25 retriever that selects the top-K relevant profiles from the user history, while Section 5.1.1 states 'We randomly sample K=2 representative profiles per user.' These are different procedures. If the reported experiments used random sampling, then the retriever component is not exercised and the method's contribution is different from the one described in Section 3.1. If BM25 retrieval was used, the experimental setup is misreported. This inconsistency needs to be resolved, and the actual selection mechanism must be stated precisely.
[Table 1] The claim that PrefReward 'consistently achieves the best performance across both models' is not supported by the reported statistics. In the zero-shot Llama-2 setting, the gain over NonPers is 0.3161 - 0.3139 = 0.0022 ROUGE-1, 0.0005 ROUGE-L, and 0.0038 METEOR. These differences are likely within run-to-run noise. No error bars, standard deviations, significance tests, or per-user paired analyses are provided. The aligned Gemma results show larger gains, but still no confidence intervals or significance testing. The paper should either provide variance estimates and statistical tests or temper the claim of consistency.
[Section 3.2, Eq. (R_i = -KL(p_y(i) || \bar P_u))] The KL divergence is defined only for probability distributions, but it is not specified whether the preference vectors v_i and p_y are normalized to sum to one. The extraction procedure averages token-level logits or filtered softmax outputs; the result may not be a valid distribution, and it may contain zeros that make KL divergent. The normalization step and the handling of zero entries must be explicitly defined. Without this, the reward value is not well-defined even if the token-to-label mapping issue were resolved.
[Sections 5.4 and Abstract] The paper claims that PrefReward improves 'generation quality and personalization interpretability', but interpretability is never evaluated. The only evidence is a single case study with one example (Figure 3). No human evaluation, no analysis of the preference matrix, no comparison of how the preference vectors differ across users, and no quantitative measure of interpretability are provided. If interpretability is a central contribution, it must be measured or at least systematically demonstrated.

minor comments (6)

[Notation, Section 2 and Section 3.1] The symbol P_u is used both for the set of profile pairs in Eq. (2) and for the preference matrix in Section 3.1. This is confusing, especially since the matrix is described as having dimension K x N while the set is a list of K pairs. Please use distinct notation.
[Figure 2] Figure 2 lacks axis labels, legend entries, and a description of which label subsets are used for each point. The recommendation that 'around 20' labels is optimal is therefore unverifiable from the figure.
[Figure 1] There is a typo: 'Consice' should be 'Concise'. The figure also does not explain what the matrix entries (e.g., 0.25, 0.07) represent, and the token-to-label mapping is not shown.
[Section 5.1.4] The 'aligned' Gemma-2B-IT setup is under-specified. The paper says the model is aligned using 'a dataset of (prompt, target output) pairs' but gives no data size, training procedure, number of steps, or learning rate. This makes the aligned condition difficult to reproduce.
[Related Work and Baselines] Drift [4] is cited as the inspiration for the preference label set and is closely related to the proposed decoding-time personalization, but it is not included as a baseline in Table 1. A comparison to this direct prior work would strengthen the evaluation.
[General] No code, model outputs, or hyperparameter configuration files are provided. Given the sensitivity of the results to sampling parameters (temperature, nucleus p, N_c, K), a complete configuration table and code release are needed for reproducibility.

Circularity Check

0 steps flagged

No circular dependency: the preference matrix and KL reward are constructed from user profiles and evaluated against external ground truth.

full rationale

The derivation chain is self-contained. In Section 3.1, the user preference matrix is built by averaging token-level LLM logits filtered to a fixed set of 41 stylistic labels; in Section 3.2, candidates are scored by -KL(p_y(i) || \bar P_u) with the user mean preference vector. This is a fixed, training-free scoring rule: no parameter is fitted to the test targets, and the reported ROUGE-1/ROUGE-L/METEOR comparisons are measured against ground-truth outputs that are not used in preference extraction. The label set is borrowed from the external Drift framework [4], not from the authors' own prior work, so the citation is not self-referential. The main caveat is that the phrase 'tokens that belong to P' is unspecified — no concrete mapping from vocabulary tokens to abstract labels such as formal or humor is provided. This is a reproducibility / missing-definition issue, not a circularity, because the preference representation and reward are not defined in terms of the evaluation labels or ground-truth outputs. No equation or fitted parameter reduces to its own input by construction, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new ontological entities. It relies on a hand-chosen label set and several architectural hyperparameters. The main assumptions concern the validity of token-level logits as a style signal and the sufficiency of averaging.

free parameters (4)

K (number of retrieved profiles) = 2
Number of user profiles used to build the preference matrix; set to 2 by hand in §5.1.1 without ablations.
N_c (number of candidate outputs) = 10
Best-of-N candidates per prompt; chosen in §5.1.4 without sensitivity analysis.
Preference label set P = 41 stylistic dimensions (e.g., formal, concise, vivid, humor)
Taken from Drift [4], not listed in full; the specific set determines the matrix dimensions and is a hand-chosen modeling choice.
temperature and nucleus p for sampling = tau=0.8, p=0.95
Sampling hyperparameters for candidate generation; chosen arbitrarily.

axioms (4)

domain assumption Token-level logits from an LLM, filtered to tokens matching stylistic labels, capture the stylistic attributes of a text.
Invoked in §3.1 'Token-Level Preference Extraction'; the entire preference matrix is built on this premise, and no evidence is given that logits over label words encode style reliably.
domain assumption The mean of per-token preference vectors across a user's profile represents the user's stable style.
Used in §3.2 to define the user preference matrix and the reward target; no validation that averaging is appropriate.
domain assumption KL divergence between preference vectors is a suitable alignment reward for selecting personalized outputs.
The reward function in §3.2 is defined as negative KL; this is a design choice with no theoretical or empirical justification.
domain assumption The LongLaMP ground-truth reviews reflect the user's personal style.
The evaluation compares against ground-truth reviews; if those are not style-consistent, the metrics do not measure personalization.

pith-pipeline@v1.3.0-alltime-deepseek · 6222 in / 10393 out tokens · 95986 ms · 2026-08-01T08:31:10.239463+00:00 · methodology

0 comments

read the original abstract

Large Language Models (LLMs) have demonstrated remarkable ability in generating personalized content by leveraging user histories and contextual cues. However, most existing personalization approaches rely on implicit representations within model parameters, making it difficult to interpret user-specific preferences or effectively handle long-context dependencies. To address these challenges, we propose PrefReward, a novel preference-aware generative framework that explicitly models user styles through a structured preference matrix and integrates it into the decoding process as a reward signal. PrefReward consists of two stages: (1) extracting a user-specific preference matrix that summarizes individual stylistic tendencies, and (2) using the matrix to guide generation via a KL-divergence-based reward function. Experiments on the LongLaMP dataset show that PrefReward outperforms non-personalized and retrieval-based baselines in both generation quality and personalization interpretability.

Figures

Figures reproduced from arXiv: 2607.21067 by Chengbing Wang, Fuli Feng, Xiaoyan Zhao, Yang Zhang, Yimeng Bai, Yue Wu.

**Figure 2.** Figure 2: Effect of different preference label size on model [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Case study of personalized generation in LongLaMP [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 6 linked inside Pith

[1]

Bashar Alhafni, Vivek Kulkarni, Dhruv Kumar, and Vipul Raheja. 2024. Per- sonalized Text Generation with Fine-Grained Linguistic Control. InProceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024), Ameet Deshpande, EunJeong Hwang, Vishvak Murahari, Joon Sung Park, Diyi Yang, Ashish Sabharwal, Karthik Narasimhan, and A...

2024
[2]

Schrum, and Anca Dragan

Jerry Zhi-Yang He, Sashrika Pandey, Mariah L. Schrum, and Anca Dragan
[3]

Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation.ACM Comput. Surv.55, 12, Article 248 (March 2023), 38 pages

2023
[4]

Minbeom Kim, Kang il Lee, Seongho Joo, Hwaran Lee, Thibaut Thonet, and Kyomin Jung. 2025. Drift: Decoding-time Personalized Alignments with Implicit User Preferences. arXiv:2502.14289

Pith/arXiv arXiv 2025
[5]

Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and Scott A. Hale. 2024. The benefits, risks and bounds of personalizing the alignment of large language models to individuals.Nature Machine Intelligence6 (2024), 383 – 392

2024
[6]

Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, Nedim Lipka, Chien Van Nguyen, Thien Huu Nguyen, and Hamed Zamani

Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A. Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, Nedim Lipka, Chien Van Nguyen, Thien Huu Nguyen, and Hamed Zamani. 2024. LongLaMP: A Benchmark for Personalized Long-form Text Generation. arXiv:2407.11016

Pith/arXiv arXiv 2024
[7]

Yibin Lei, Liang Ding, Yu Cao, Changtong Zan, Andrew Yates, and Dacheng Tao. 2023. Unsupervised Dense Retrieval with Relevance-Aware Contrastive Pre-Training. InFindings of the Association for Computational Linguistics: ACL 2023, Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 10932–10940

2023
[8]

Sheshera Mysore, Zhuoran Lu, Mengting Wan, Longqi Yang, Steve Menezes, Tina Baghaee, Emmanuel Barajas Gonzalez, Jennifer Neville, and Tara Safavi. 2023. Pearl: Personalizing Large Language Model Writing Assistants with Generation- Calibrated Retrievers.ArXivabs/2311.09180 (2023)

Pith/arXiv arXiv 2023
[9]

Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Frame- work: BM25 and Beyond.Found. Trends Inf. Retr.3, 4 (April 2009), 333–389

2009
[10]

Alireza Salemi, Surya Kallumadi, and Hamed Zamani. 2024. Optimization Meth- ods for Personalizing Large Language Models through Retrieval Augmentation. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval(2024)

2024
[11]

Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thail...

2024
[12]

Jinghao Zhang, Yuting Liu, Wenjie Wang, Qiang Liu, Shu Wu, Liang Wang, and Tat-Seng Chua. 2025. Personalized Text Generation with Contrastive Activation Steering. InProceedings of the 63rd Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (...

2025
[13]

Zhaowei Zhang, Fengshuo Bai, Qizhi Chen, Chengdong Ma, Mingzhi Wang, Hao- ran Sun, Zilong Zheng, and Yaodong Yang. 2025. Amulet: ReAlignment During Test Time for Personalized Preference Adaptation of LLMs. arXiv:2502.19148

Pith/arXiv arXiv 2025
[14]

Zhehao Zhang, Ryan A. Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, Ruiyi Zhang, Jiuxiang Gu, Tyler Derr, Hongjie Chen, Junda Wu, Xiang Chen, Zichao Wang, Subrata Mitra, Nedim Lipka, Nesreen Ahmed, and Yu Wang. 2025. Personalization of Large Language Models: A Survey. arXiv:2411.00027

Pith/arXiv arXiv 2025
[2025]

arXiv:2405.01768

Context Steering: Controllable Personalization at Inference Time. arXiv:2405.01768

Pith/arXiv arXiv

[1] [1]

Bashar Alhafni, Vivek Kulkarni, Dhruv Kumar, and Vipul Raheja. 2024. Per- sonalized Text Generation with Fine-Grained Linguistic Control. InProceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024), Ameet Deshpande, EunJeong Hwang, Vishvak Murahari, Joon Sung Park, Diyi Yang, Ashish Sabharwal, Karthik Narasimhan, and A...

2024

[2] [2]

Schrum, and Anca Dragan

Jerry Zhi-Yang He, Sashrika Pandey, Mariah L. Schrum, and Anca Dragan

[3] [3]

Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation.ACM Comput. Surv.55, 12, Article 248 (March 2023), 38 pages

2023

[4] [4]

Minbeom Kim, Kang il Lee, Seongho Joo, Hwaran Lee, Thibaut Thonet, and Kyomin Jung. 2025. Drift: Decoding-time Personalized Alignments with Implicit User Preferences. arXiv:2502.14289

Pith/arXiv arXiv 2025

[5] [5]

Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and Scott A. Hale. 2024. The benefits, risks and bounds of personalizing the alignment of large language models to individuals.Nature Machine Intelligence6 (2024), 383 – 392

2024

[6] [6]

Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, Nedim Lipka, Chien Van Nguyen, Thien Huu Nguyen, and Hamed Zamani

Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A. Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, Nedim Lipka, Chien Van Nguyen, Thien Huu Nguyen, and Hamed Zamani. 2024. LongLaMP: A Benchmark for Personalized Long-form Text Generation. arXiv:2407.11016

Pith/arXiv arXiv 2024

[7] [7]

Yibin Lei, Liang Ding, Yu Cao, Changtong Zan, Andrew Yates, and Dacheng Tao. 2023. Unsupervised Dense Retrieval with Relevance-Aware Contrastive Pre-Training. InFindings of the Association for Computational Linguistics: ACL 2023, Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 10932–10940

2023

[8] [8]

Sheshera Mysore, Zhuoran Lu, Mengting Wan, Longqi Yang, Steve Menezes, Tina Baghaee, Emmanuel Barajas Gonzalez, Jennifer Neville, and Tara Safavi. 2023. Pearl: Personalizing Large Language Model Writing Assistants with Generation- Calibrated Retrievers.ArXivabs/2311.09180 (2023)

Pith/arXiv arXiv 2023

[9] [9]

Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Frame- work: BM25 and Beyond.Found. Trends Inf. Retr.3, 4 (April 2009), 333–389

2009

[10] [10]

Alireza Salemi, Surya Kallumadi, and Hamed Zamani. 2024. Optimization Meth- ods for Personalizing Large Language Models through Retrieval Augmentation. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval(2024)

2024

[11] [11]

Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thail...

2024

[12] [12]

Jinghao Zhang, Yuting Liu, Wenjie Wang, Qiang Liu, Shu Wu, Liang Wang, and Tat-Seng Chua. 2025. Personalized Text Generation with Contrastive Activation Steering. InProceedings of the 63rd Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (...

2025

[13] [13]

Zhaowei Zhang, Fengshuo Bai, Qizhi Chen, Chengdong Ma, Mingzhi Wang, Hao- ran Sun, Zilong Zheng, and Yaodong Yang. 2025. Amulet: ReAlignment During Test Time for Personalized Preference Adaptation of LLMs. arXiv:2502.19148

Pith/arXiv arXiv 2025

[14] [14]

Zhehao Zhang, Ryan A. Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, Ruiyi Zhang, Jiuxiang Gu, Tyler Derr, Hongjie Chen, Junda Wu, Xiang Chen, Zichao Wang, Subrata Mitra, Nedim Lipka, Nesreen Ahmed, and Yu Wang. 2025. Personalization of Large Language Models: A Survey. arXiv:2411.00027

Pith/arXiv arXiv 2025

[15] [2025]

arXiv:2405.01768

Context Steering: Controllable Personalization at Inference Time. arXiv:2405.01768

Pith/arXiv arXiv