Pith. sign in

REVIEW 2 major objections 5 minor 62 references

Learning Preference Adaptation for Large Language Model Personalization via Verbal Reinforcement Learning

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A learned text policy can rewrite a user's universal preference profile into a short task-specific one, improving 33 of 39 task-model cells at 22.8 percent of the original length.

desk verdict A clean, well-scoped personalization method whose reported gains are probably real but partly unattributed, because the paper never runs the T=0 control. read the letter →

arxiv 2608.09507 v2 pith:MEF4URI7 submitted 2026-08-10 cs.CL cs.AI

Yuting Liu , Wei Wu , Jianzhe Zhao , Guibing Guo This is my paper · ORCID
classification cs.CLcs.AI
keywords largelanguagemodelpersonalizationtask-specificpreferenceadaptationverbalreinforcementlearningmeta-learninguserprofilescontextcompressionfaithfulnessaudit
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large language models often personalize by stuffing a user's complete natural-language preference summary into the prompt. This paper argues that such universal summaries are wasteful and even harmful, because most of their content is irrelevant to any single task, and proposes AlignXada, a training-free meta-learning framework that learns a short textual rewrite rule per task. The rule is induced from a small support set of demonstrations through verbal reinforcement learning and then applied by a frozen refiner to turn the universal profile into a task-adapted one. On 39 task-model combinations the rewritten profiles improve 33 of them by an average of 3.82 points over the raw profile, keep only 22.8 percent of the original tokens, and beat a retrieval-based baseline in 36 cells. A faithfulness audit reports that 97.5 percent of refined-profile claims are supported by the source profiles, so the gains come from reorganization rather than hallucination.

What carries the argument

The load-bearing mechanism is the task-specific textual refinement policy $\phi^{(\tau)}$: a short natural-language instruction that tells a frozen refiner how to compress the universal profile, what evidence to preserve, what to avoid, and how to order priorities. The policy is iteratively revised by a frozen meta-learner through verbal reinforcement learning, meaning rollout of the current policy on a support set, aggregation of the resulting scores and failure patterns into structured text feedback, and a prompt-based update that emits the next policy. The adapted profile $\tilde{P}_u = R(P_u, \phi^{(\tau)})$ is what the downstream model actually conditions on, so the policy, not the model weights, carries all of the task adaptation. This design decouples policy generation from preference refinement and makes the adaptation auditable as text.

What would settle it

Run the induction procedure on a task whose support set is deliberately built from a single narrow topic or user group, then evaluate on held-out users outside that group; if the refined profile no longer beats the raw universal profile, the representativeness assumption has failed. A complementary check is to use tasks whose universal profiles contain no task-relevant evidence at all, where the method should show no gain because a refiner can only retain what the source profile already contains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that the best prompt for personalization is not the user's universal preference summary but a task-conditioned rewrite of it. The framework defines the adapted profile as $\tilde{P}_u = R(P_u, \phi^{(\tau)})$, where $R$ is a frozen refiner and $\phi^{(\tau)}$ is a natural-language refinement policy learned by a frozen meta-learner. The meta-learner iterates rollout, structured feedback aggregation, and policy revision over a support set $S^{(\tau)} = \{(P_{u_i}, x_i, y_i)\}$, optimizing the downstream score $\mathrm{Score}_\tau(M(x_i, \tilde{P}_{u_i}), y_i)$; the best policy on a disjoint development set is then deployed. The empirical core is a 39-cell evaluation in which 33 cells improve, the average gain is $+3.82$ points, the refined profiles retain $22.8\%$ of the original tokens, and a retrieval-based baseline is outperformed in 36 cells. The faithfulness audit reports $97.5\%$ of refined-profile claims are supported by the source profile, with $83.3\%$ conditional retention of available gold preference evidence; the paper reads the main bottleneck as incomplete source coverage ($34.0\%$) rather than refiner hallucination.

Load-bearing premise

The load-bearing premise is that the small set of example tasks used to learn the rewrite policy represents the real distribution of tasks and users; if those examples are biased or narrow, the learned rewrite rule will not transfer to held-out users.

Editorial extensions

If this is right

  • If the claim holds, a lifelong agent can maintain one universal memory per user and derive compact task-specific profiles on demand, avoiding the need to store or retrieve separate memories for every task.
  • Downstream accuracy improves while context shrinks to roughly a fifth of the original, and the near-zero correlation between gain and token ratio ($r=0.06$) indicates that gains come from retaining the right evidence rather than keeping more text.
  • Task-oriented profile rewriting beats a retrieval-based baseline in 36 of 39 cells, suggesting that consolidated preference reorganization can outperform locally retrieved excerpts for personalization.
  • The benefit persists on cleaner source-native profiles, with average gains of $+1.30$ and $+1.69$ points, so it is not just the removal of cross-domain noise from a merged benchmark.
  • Policies can be induced from a small support set ($b=20$) and usually converge within a few update rounds, which makes the approach usable with proprietary models because no parameters are trained.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the policy is plain text, users or system builders could inspect and edit the rewrite rule before deployment, making the adaptation auditable; the paper emphasizes transparency but does not itself test human editing.
  • Editorial inference: if support-set representativeness is the binding constraint, a natural next step is to learn policies from more diverse support sets and to test whether a policy transfers across task families; the paper's stated limitation points exactly here.
  • Editorial inference: the high conditional retention of available evidence suggests headroom lies upstream in universal profile construction, so improving source coverage may matter more than further compression; the paper identifies this as a bottleneck rather than a tested improvement.
  • Editorial inference: because gains are uncorrelated with token ratio in the tested range, one could push compression further to find where performance breaks; this is a testable extension, not a claim of the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper introduces AlignXada, a training-free meta-learning framework that learns a reusable natural-language refinement policy via verbal reinforcement learning. Given a universal user preference summary P_u and a downstream task, AlignXada induces a task-specific policy from a small support set of demonstrations and uses a frozen refiner to produce a compact, task-adapted profile. The policy is iteratively revised by a frozen meta learner using structured feedback, and the best policy is selected on a disjoint development set before held-out deployment. The work is evaluated on a composite benchmark built from PersonaMem-v2 and MemoryCD across 13 tasks and three downstream models, reporting an average gain of +3.82 points over the raw universal profile, a token ratio of 22.8%, and superiority over a BM25 RAG baseline in 36 of 39 cells. The paper also reports results on the two source-native benchmarks, ablations of support batch size, update rounds, adaptive sampling, meta-model generalization, preference-source robustness, and a claim-level faithfulness and evidence-retention audit.

Significance. If the central claim holds, AlignXada is a practical and interpretable complement to universal profile construction for lifelong personalization: it converts a single task-agnostic preference summary into task-specific views without parameter updates, is compatible with proprietary models, and the faithfulness audit suggests the process is controlled compression rather than hallucination. The paper has several genuine strengths: user-disjoint support, development, and evaluation splits; policy selection on a disjoint development set; an additional source-native evaluation that checks robustness to cross-domain noise introduced by benchmark construction; ablations of batch size, update rounds, adaptive sampling, and meta-model family; a faithfulness audit with explicit metrics; and release of code. The main empirical gap is that no T=0 (initial-policy-only) condition is reported, which is needed to attribute the gains to verbal reinforcement learning rather than to the initial compression instruction, and the headline results are single-run point estimates without variance or confidence intervals.

major comments (2)
  1. [§4.2, Algorithm 1, Appendix F.1] The paper never reports a deployment baseline that uses only the initial task-agnostic policy φ0 described in Appendix F.1. Because φ0 is already a compression instruction ("Rewrite the profile into concise, stable preference rules that preserve relevant user signals while reducing unnecessary detail"), the reported +3.82 average gain and 22.8% token ratio could in principle be produced by φ0 alone, with the verbal-RL rounds contributing little or nothing. Figure 4 makes this gap concrete: for Professional Writing the final selected policy is exactly the Initial policy, so the RL loop is not needed for that cell. The batch-size ablation in Figure 3 provides indirect evidence that the update procedure matters, but it does not isolate φ0. I request a T=0 condition that runs the deployment stage (Eqs. 3–5) with φ0 only, on the same evaluation sets, reporting both the primary metric and token ratio; this is a minimal control that directly tests the "via verbal reinforcement learning" attribution in the title and abstract.
  2. [§4.1–4.2, Table 1] All headline numbers are single-run point estimates with no variance or confidence intervals. Table 3 shows evaluation query counts as low as 84–131 for several tasks, and many reported deltas are small (e.g., +0.43 for Review Generation with GPT-5-mini and -0.67 for Professional Email with Qwen3-8B), so the "improves 33 of 39 cells" claim may partly reflect noise. The exact sign test reported in the Table 1 caption aggregates task-level gains and does not quantify within-cell uncertainty. I request bootstrap confidence intervals or multiple support-set/run replications, at least for the composite benchmark's headline comparison.
minor comments (5)
  1. [§4.4, Figure 4] The caption of Figure 4 should state explicitly what the "Initial" bar means — that the final selected policy for Professional Writing is φ0 itself — because this is an important observation for interpreting the role of the update rounds.
  2. [§3.1, Eq. (2)] The expectation in Eq. (2) is written over S(τ) but no sampling distribution is defined; please clarify whether this is an empirical average over the support set or over a stochastic sampling process.
  3. [§4.1, RAG baseline description] The sentence "We use the downstream task prompt as the retrieval query" is ambiguous: please specify whether the retrieval query is the user input x_i, a task-level description, or a concatenation of both, since this affects the interpretation of the RAG comparison.
  4. [Appendix D.2] The faithfulness audit relies entirely on labels produced by DeepSeek-V4-Pro as the judge model, with no reported inter-annotator agreement or human validation; a small human-annotated sample would substantially strengthen the audit's reliability.
  5. [Appendix C] Please state explicitly whether the all-DeepSeek variant AlignXada-D uses the same number of update rounds T=5 and the same adaptive sampling configuration as the main experiments, so that the comparison isolates the meta-model change.

Circularity Check

0 steps flagged · score 0.0 of 10

No construction-level circularity: the refinement policy is induced on a support set, selected on a disjoint development set, and the reported gains are measured on user-disjoint held-out queries.

full rationale

AlignXada's derivation chain is externally grounded rather than self-referential. The universal profile P_u is an input; the refiner applies policy phi_t to produce P~_u = R(P_u, phi_t) (Eq. 3); the downstream model M yields y-hat (Eq. 4); and the policy is updated from support-set feedback (Eqs. 9-10) and selected by development-set utility (Eqs. 7, 11). The paper states: 'For each task, we construct user-disjoint support, development, and evaluation sets, ensuring that users involved in policy induction do not appear in the held-out evaluation set.' The headline +3.82 average gain and 22.8% token ratio are therefore measured on held-out queries, not on the fitted support or development examples, so no fitted parameter is renamed as a prediction. The self-citations (CoRA and Text-as-Universal-Interface) appear only as background in the introduction and related work and are not the source of the benchmark, the downstream models, or the evaluation protocol, so they are not load-bearing. The closest concerns are a missing T=0 ablation that would isolate the contribution of the verbal-RL updates from the initial compression instruction phi_0, and possible selection of hyperparameters (b=20, T=5) on the composite benchmark; these are correctness and overfitting risks rather than circular reductions, and the paper itself flags the broader generalization assumption in Limitations: 'AlignXada assumes that the support set used for policy induction is representative of the target task distribution. When the support set is biased or insufficiently diverse, the induced refinement policy may not generalize well.' No equation in the paper defines the target result in terms of itself.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's central claim rests on a small number of domain assumptions: the availability of a universal profile, the representativeness of support sets, the validity of task-specific scores as rewards, and the reliability of the judge model. The free parameters are hyperparameters and the handcrafted initial policy; the numbers most directly affecting the headline result are b=20 and T=5. No invented entities are introduced.

free parameters (5)
  • Support batch size b = 20
    Selected in Section 4.4, Figure 3 as the best accuracy-compression trade-off among 5, 10, 20; all main results use b=20.
  • Number of update rounds T = 5
    Selected in Section 4.4, Figure 4 as a balance between policy quality and inference cost; some tasks (e.g., rating prediction) only find their best policy after round 5.
  • RAG retrieval window size = 120 words with 30-word overlap, top 8 chunks
    These parameters define the RAG baseline used for the claim 'outperforms RAG in 36 cells'; different chunking could alter the baseline strength.
  • RAG BM25 parameters = k1=1.5, b=0.75
    Standard BM25 defaults chosen for the RAG baseline that AlignXada is compared against.
  • Initial policy text = 'Rewrite the profile into concise, stable preference rules...'
    Handcrafted task-agnostic starting policy (Appendix F.1); the policy induction starts from this text, so it shapes the induced policies.
assumptions (5)
  • domain assumption A universal preference summary P_u is available for each user and treated as prior knowledge.
    Section 3.1: 'let P_u denote the corresponding universal preference summary, which may be obtained from an external LLM or memory system and is treated as prior knowledge in this work.' The method does not construct or verify this profile.
  • domain assumption The support set S(τ) is representative of the target task distribution.
    Section 5 Limitations: 'AlignXada assumes that the support set used for policy induction is representative of the target task distribution. When the support set is biased or insufficiently diverse, the induced refinement policy may not generalize well.'
  • domain assumption The task-specific score Scoreτ on support and development examples is a valid reward signal for personalization quality.
    Section 3.1 Eq. (2) uses expected score as the optimization objective; the policy is selected by development-set scores (Eq. 11). No human validation of the scores is reported.
  • standard math The development set D(τ) is disjoint from support and evaluation sets and predicts held-out performance.
    Section 3.1 and Algorithm 1: policy selection uses the development set; the paper states user-disjoint splits, so this is a standard ML assumption rather than an ad hoc one.
  • domain assumption DeepSeek-V4-Pro judge labels for claim support and evidence retention are reliable.
    Appendix D.2: the faithfulness audit uses DeepSeek-V4-Pro as the judge with no reported inter-annotator agreement or human validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Preference Adaptation for Large Language Model Personalization via Verbal Reinforcement Learning." pith.science (2026). https://pith.science/paper/MEF4URI7

@misc{pith2026260809507,
  author       = {Pith},
  title        = {Pith review of: Learning Preference Adaptation for Large Language Model Personalization via Verbal Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEF4URI7}},
  note         = {Machine review of arXiv:2608.09507}
}
read the original abstract

Natural language user preferences provide an interpretable interface for LLM personalization. However, universal preference summaries often contain information irrelevant to a particular downstream task. Directly supplying the full preference summary therefore wastes context capacity and introduces cross-task distraction, while manually designing task-specific preference views is difficult to scale. In this work, we study \emph{task-specific preference adaptation}: given a universal user preference summary and a downstream task, derive a task-conditioned representation that preserves sufficient decision-relevant evidence while removing redundant context. To this end, we propose \textsc{AlignXada}, a training-free meta-learning framework that induces reusable textual refinement policies for adapting universal preference summaries to task-specific ones. The refinement policy is iteratively optimized by a meta learner through verbal reinforcement learning. Across 13 tasks and three downstream models (39 task--model cells), \textsc{AlignXada} achieves an average gain of 3.82 points, improving 33 cells while retaining only 22.8\% of the original profile tokens and outperforming RAG in 36 cells. An extended faithfulness analysis further shows that the refined profiles remain largely grounded in the source preferences while preserving task-relevant personalization signals, suggesting that profile-side adaptation serves as a practical complement to universal memory construction for lifelong personalized agents.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 10 linked inside Pith

  1. [1]

    2025 , eprint =

    Jiang, Bowen and Yuan, Yuan and Shen, Maohao and Hao, Zhuoqun and Xu, Zhangchen and Chen, Zichen and Liu, Ziyi and Vijjini, Anvesh Rao and He, Jiashu and Yu, Hanchao and Poovendran, Radha and Wornell, Gregory and Ungar, Lyle and Roth, Dan and Chen, Sihao and Taylor, Camillo Jose , journal =. 2025 , eprint =

    2025

  2. [2]

    2024 , url =

    Salemi, Alireza and Mysore, Sheshera and Bendersky, Michael and Zamani, Hamed , booktitle =. 2024 , url =

    2024

  3. [3]

    Kumar, Ishita and Viswanathan, Snigdha and Yerra, Sushrita and Salemi, Alireza and Rossi, Ryan A. and Dernoncourt, Franck and Deilamsalehy, Hanieh and Chen, Xiang and Zhang, Ruiyi and Agarwal, Shubham and Lipka, Nedim and Nguyen, Chien Van and Nguyen, Thien Huu and Zamani, Hamed , journal =. 2024 , eprint =

    2024

  4. [4]

    Optimizing User Profiles via Contextual Bandits for Retrieval-Augmented

    Du, Linfeng and Yuan, Ye and Zhao, Zichen and Lyu, Fuyuan and Penaloza, Emiliano and Chen, Xiuying and Sun, Zipeng and Kang, Jikun and Charlin, Laurent and Liu, Xue and Wu, Haolun , year =. Optimizing User Profiles via Contextual Bandits for Retrieval-Augmented. 2601.12078 , archivePrefix =

    arXiv

  5. [5]

    Persona-

    Sun, Chenkai and Yang, Ke and Gangi Reddy, Revanth and Fung, Yi and Chan, Hou Pong and Small, Kevin and Zhai, ChengXiang and Ji, Heng , booktitle =. Persona-. 2025 , url =

    2025

  6. [6]

    2023 , url =

    Liu, Shuai and Cho, Hyundong and Freedman, Marjorie and Ma, Xuezhe and May, Jonathan , booktitle =. 2023 , url =

    2023

  7. [7]

    arXiv preprint arXiv:2504.05731 , year =

    Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation , author =. arXiv preprint arXiv:2504.05731 , year =. 2504.05731 , archivePrefix =

    arXiv

  8. [8]

    arXiv preprint arXiv:2402.04401 , year =

    Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning , author =. arXiv preprint arXiv:2402.04401 , year =. 2402.04401 , archivePrefix =

    arXiv

Show all 62 references
  1. [9]

    2024 , eprint =

    Clarke, Christopher and Heng, Yuzhao and Tang, Lingjia and Mars, Jason , journal =. 2024 , eprint =

    2024

  2. [10]

    ACM Transactions on Information Systems , volume =

    Personalized Prompt Learning for Explainable Recommendation , author =. ACM Transactions on Information Systems , volume =. 2023 , doi =

    2023

  3. [11]

    2025 , url =

    Liu, Jiongnan and Zhu, Yutao and Wang, Shuting and Wei, Xiaochi and Min, Erxue and Lu, Yu and Wang, Shuaiqiang and Yin, Dawei and Dou, Zhicheng , booktitle =. 2025 , url =

    2025

  4. [12]

    arXiv preprint arXiv:2310.11564 , year =

    Personalized Soups: Personalized Large Language Model Alignment via Post-hoc Parameter Merging , author =. arXiv preprint arXiv:2310.11564 , year =. 2310.11564 , archivePrefix =

    arXiv

  5. [13]

    2024 , note =

    Zhuang, Yuchen and Sun, Haotian and Yu, Yue and Qiang, Rushi and Wang, Qifan and Zhang, Chao and Dai, Bo , booktitle =. 2024 , note =

    2024

  6. [14]

    arXiv preprint arXiv:2310.20081 , year =

    Integrating Summarization and Retrieval for Enhanced Personalization via Large Language Models , author =. arXiv preprint arXiv:2310.20081 , year =. 2310.20081 , archivePrefix =

    arXiv

  7. [15]

    Proceedings of the 1st Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U) , pages =

    Pearl: Personalizing Large Language Model Writing Assistants with Generation-Calibrated Retrievers , author =. Proceedings of the 1st Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U) , pages ...

    2024

  8. [16]

    arXiv preprint arXiv:2601.04963 , year =

    Text as a Universal Interface for Transferable Personalization , author =. arXiv preprint arXiv:2601.04963 , year =. 2601.04963 , archivePrefix =

  9. [17]

    International Conference on Learning Representations (ICLR) , year =

    Large Language Models as Optimizers , author =. International Conference on Learning Representations (ICLR) , year =

  10. [18]

    2024 , url =

    Guo, Qingyan and Wang, Rui and Guo, Junliang and Li, Bei and Song, Kaitao and Tan, Xu and Liu, Guoqing and Bian, Jiang and Yang, Yujiu , booktitle =. 2024 , url =

    2024

  11. [19]

    2024 , eprint =

    Yuksekgonul, Mert and Bianchi, Federico and Boen, Joseph and Liu, Sheng and Huang, Zhi and Guestrin, Carlos and Zou, James , journal =. 2024 , eprint =

    2024

  12. [20]

    Advances in Neural Information Processing Systems (NeurIPS) , pages =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , pages =. 2023 , url =

    2023

  13. [21]

    and Stoica, Ion and Gonzalez, Joseph E

    Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , journal =. 2023 , eprint =

    2023

  14. [22]

    2024 , doi =

    Zhong, Wanjun and Guo, Lianghong and Gao, Qiqi and Ye, He and Wang, Yanlin , booktitle =. 2024 , doi =

    2024

  15. [23]

    2025 , note =

    Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng , booktitle =. 2025 , note =

    2025

  16. [24]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Augmenting Language Models with Long-Term Memory , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  17. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Language Models are Few-Shot Learners , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  18. [26]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  19. [27]

    International Conference on Machine Learning (ICML) , pages =

    Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks , author =. International Conference on Machine Learning (ICML) , pages =. 2017 , url =

    2017

  20. [28]

    2025 , eprint =

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities , author =. 2025 , eprint =

    2025

  21. [29]

    2026 , month = feb, howpublished =

    Gemini 3.1 Pro Model Card , author =. 2026 , month = feb, howpublished =

    2026

  22. [30]

    2025 , eprint =

    Qwen3 Technical Report , author =. 2025 , eprint =

    2025

  23. [31]

    Grattafiori, Aaron and others , year =. The. 2407.21783 , archivePrefix =

    arXiv

  24. [32]

    2023 , url =

    Jiang, Huiqiang and Wu, Qianhui and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , booktitle =. 2023 , url =

    2023

  25. [33]

    2024 , url =

    Jiang, Huiqiang and Wu, Qianhui and Luo, Xufang and Li, Dongsheng and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , booktitle =. 2024 , url =

    2024

  26. [34]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages =

    Pan, Zhuoshi and Wu, Qianhui and Jiang, Huiqiang and Xia, Menglin and Luo, Xufang and Zhang, Jue and Lin, Qingwei and R. Findings of the Association for Computational Linguistics: ACL 2024 , pages =. 2024 , url =

    2024

  27. [35]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Compressing Context to Enhance Inference Efficiency of Large Language Models , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =. 2023 , note =

    2023

  28. [36]

    Transactions of the Association for Computational Linguistics , year =

    Lost in the Middle: How Language Models Use Long Contexts , author =. Transactions of the Association for Computational Linguistics , year =

  29. [37]

    Training language models to follow instructions with human feedback , url =

    Ouyang, Long and Wu, Jeffrey and Jiang, Xu and Almeida, Diogo and Wainwright, Carroll and Mishkin, Pamela and Zhang, Chong and Agarwal, Sandhini and Slama, Katarina and Ray, Alex and Schulman, John and Hilton, Jacob and Kelton, Fraser and Miller, Luke and Simens, Maddie and As...

  30. [38]

    2026 , url =

    OpenClaw , title =. 2026 , url =

    2026

  31. [39]

    2026 , url =

    Hermes , title =. 2026 , url =

    2026

  32. [40]

    Proceedings of the ACM Web Conference 2024 , pages =

    Knowledge-Augmented Large Language Models for Personalized Contextual Query Suggestion , author =. Proceedings of the ACM Web Conference 2024 , pages =. 2024 , doi =

    2024

  33. [41]

    Recommendation as Language Processing (

    Geng, Shijie and Liu, Shuchang and Fu, Zuohui and Ge, Yingqiang and Zhang, Yongfeng , booktitle =. Recommendation as Language Processing (. 2022 , doi =

    2022

  34. [42]

    2023 , doi =

    Bao, Keqin and Zhang, Jizhi and Zhang, Yang and Wang, Wenjie and Feng, Fuli and He, Xiangnan , booktitle =. 2023 , doi =

    2023

  35. [43]

    2024 , publisher =

    Lyu, Hanjia and Jiang, Song and Zeng, Hanqing and Xia, Yinglong and Wang, Qifan and Zhang, Si and Chen, Ren and Leung, Chris and Tang, Jiajie and Luo, Jiebo , booktitle =. 2024 , publisher =

    2024

  36. [44]

    arXiv preprint arXiv:2409.00887 , year =

    User-Specific Dialogue Generation with User Profile-Aware Pre-Training Model and Parameter-Efficient Fine-Tuning , author =. arXiv preprint arXiv:2409.00887 , year =. 2409.00887 , archivePrefix =

    arXiv

  37. [45]

    2025 , doi =

    Liu, Yuting and Zhang, Jinghao and Dang, Yizhou and Liang, Yuliang and Liu, Qiang and Guo, Guibing and Zhao, Jianzhe and Wang, Xingwei , booktitle =. 2025 , doi =

    2025

  38. [46]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Personalized Text Generation with Contrastive Activation Steering , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2025 , url =

    2025

  39. [47]

    arXiv preprint arXiv:2404.05970 , year =

    Optimization Methods for Personalizing Large Language Models through Retrieval Augmentation , author =. arXiv preprint arXiv:2404.05970 , year =. 2404.05970 , archivePrefix =

    arXiv

  40. [48]

    arXiv preprint arXiv:2507.13579 , year =

    Learning to Summarize User Information for Personalized Reinforcement Learning from Human Feedback , author =. arXiv preprint arXiv:2507.13579 , year =. 2507.13579 , archivePrefix =

  41. [49]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =

    Guided Profile Generation Improves Personalization with Large Language Models , author =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =. 2024 , publisher =

    2024

  42. [50]

    2025 , eprint =

    Chen, Yizhuo and Liu, Xin and Wang, Ruijie and Li, Zheng and Chen, Pei and Yu, Changlong and Yin, Qingyu and Nigam, Priyanka and Jiang, Meng and Yin, Bing , journal =. 2025 , eprint =

    2025

  43. [51]

    Computer , volume =

    Matrix Factorization Techniques for Recommender Systems , author =. Computer , volume =. 2009 , doi =

    2009

  44. [52]

    2026 , eprint=

    SensorPersona: An LLM-Empowered System for Continual Persona Extraction from Longitudinal Mobile Sensor Streams , author=. 2026 , eprint=

    2026

  45. [53]

    2026 , eprint=

    MemRerank: Preference Memory for Personalized Product Reranking , author=. 2026 , eprint=

    2026

  46. [54]

    arXiv preprint arXiv:2504.19413 , year=

    Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory , author=. arXiv preprint arXiv:2504.19413 , year=

    arXiv

  47. [55]

    arXiv preprint arXiv:2507.02259 , year=

    MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent , author=. arXiv preprint arXiv:2507.02259 , year=

    arXiv

  48. [56]

    2023 , booktitle =

    SteerLM: Attribute Conditioned SFT as an (User-Steerable) Alternative to RLHF , author=. 2023 , booktitle =

    2023

  49. [57]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Rewards-in-Context: Multi-objective Alignment of Foundation Models with Dynamic Preference Adjustment , author =. Proceedings of the 41st International Conference on Machine Learning , pages =

  50. [58]

    ``In-Dialogues We Learn'': Towards Personalized Dialogue Without Pre-defined Profiles through In-Dialogue Learning

    Cheng, Chuanqi and Tu, Quan and Wu, Wei and Shang, Shuo and Mao, Cunli and Yu, Zhengtao and Yan, Rui. ``In-Dialogues We Learn'': Towards Personalized Dialogue Without Pre-defined Profiles through In-Dialogue Learning. Proceedings of the 2024 Conference on Empirical Methods in ...

    2024

  51. [59]

    ArXiv , year=

    From 1,000,000 Users to Every User: Scaling Up Personalized Preference for User-level Alignment , author=. ArXiv , year=

  52. [60]

    ArXiv , year=

    Extended Inductive Reasoning for Personalized Preference Inference from Behavioral Signals , author=. ArXiv , year=

  53. [61]

    arXiv preprint arXiv:2603.25973 , year=

    Memorycd: Benchmarking long-context user memory of llm agents for lifelong cross-domain personalization , author=. arXiv preprint arXiv:2603.25973 , year=

  54. [62]

    2009 , journal =

    Robertson, Stephen and Zaragoza, Hugo , title =. 2009 , journal =

    2009

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.