Pith. sign in

REVIEW 4 major objections 5 minor 58 references

SPARK claims that personalized search can be built from many persona-specialized LLM agents coordinated by a context-aware router and layered memory, replacing the single static user profile.

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

2026-08-03 13:27 UTC pith:FH7Q7M7D

load-bearing objection A clearly written framework proposal for persona-based multi-agent personalized search, but it is an architecture sketch with zero experiments and a bibliography that needs cleanup before I'd trust its scholarship. the 4 major comments →

arxiv 2512.24008 v3 pith:FH7Q7M7D submitted 2025-12-30 cs.AI

SPARK: Search Personalization via Agent-Driven Retrieval and Knowledge-sharing

Gaurab Chhetri , Subasish Das , Tausif Islam Chowdhury This is my paper
classification cs.AI
keywords search personalizationmulti-agent LLM systemspersona modelingretrieval-augmented generationcontextual bandit routingcognitive memory architectureintent-aware diversityfilter bubbles
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 argues that current personalized search is too rigid—static user profiles and one-size-fits-all ranking pipelines fail when a person's task shifts or their needs are multifaceted. SPARK proposes instead to decompose retrieval among multiple specialized LLM agents, each defined by a persona tuple of role, expertise, task context, and domain. A Persona Coordinator reads each query and its session context, activates the most relevant agents, and selects a coordination protocol: independent parallel retrieval, sequential relay, or bounded debate. Each agent runs its own retrieval-augmented generation loop with separate working, episodic, and semantic memory stores, and an arbiter fuses the ranked lists into a single grounded answer. The paper's central claim is that personalization emerges from these minimal coordination rules rather than from a hand-built personalization pipeline, and it offers concrete, testable hypotheses about when multi-agent coordination helps, how fast adaptive routing can track task drift, and how memory separation affects efficiency and diversity.

Core claim

SPARK reconceptualizes search personalization as a multi-agent coordination problem. Its central claim is that a formal persona space—quadruples of role, expertise, task context, and domain—combined with a Persona Coordinator that routes queries by a stochastic softmax over persona embeddings, produces retrieval that adapts to the current task and session rather than to a user's averaged long-term profile. Agents maintain separate working, episodic, and semantic memory and execute a retrieval-augmented generation loop; an arbiter fuses their rankings and synthesizes an evidence-backed answer. The authors assert that personalization emerges from distributed agent behavior under minimal coordi

What carries the argument

The load-bearing mechanism is the Persona Coordinator: a learned routing policy, implemented as a contextual bandit, that computes a softmax distribution over persona embeddings for each incoming query and session context, activates the top-k agents under a budget, and selects one of three coordination protocols—independent execution, relay, or constrained debate. The second key component is the tripartite memory subsystem (working, episodic, semantic), which lets agents separate transient session state from durable user preferences. The third is arbiter-side fusion, using Reciprocal Rank Fusion (rank aggregation by summed reciprocal ranks) and optionally an intent-aware diversity objective

Load-bearing premise

The load-bearing assumption is that the Persona Coordinator's learned routing policy—a contextual bandit trained on implicit feedback such as clicks and dwell time—can reliably select the right agents and protocol for each query; if this mapping is wrong, the multi-agent machinery only adds latency and cost.

What would settle it

Run the proposed offline evaluation on a session-based retrieval benchmark with logged queries, clicks, and dwell times, comparing bandit-routed SPARK against static routing and a single-agent RAG baseline. If bandit routing fails to match or exceed the static baseline within the predicted 3–5 interactions under task drift, or if the independent-specialist protocol does not outperform debate on low-complexity queries as H1 predicts, the central claims of adaptive routing and protocol selection are refuted.

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

If this is right

  • If SPARK's claim holds, a search engine does not need a single monolithic user profile; personalization can emerge at query time by activating the few persona agents whose role, expertise, and domain match the current task, making the system responsive to atypical and task-shifting queries.
  • The coordinator's contextual-bandit routing, if it learns reliable query-to-agent mappings, would let the same architecture serve both simple and complex queries: cheap parallel retrieval for easy cases, relay or debate only when predicted difficulty and ambiguity justify the cost.
  • The explicit split of working, episodic, and semantic memory predicts measurable efficiency gains—less context bloat and lower latency—without losing grounding quality, a prediction tied to cognitive-architecture principles.
  • Adding an intent-aware diversity objective to fusion should counteract filter-bubble narrowing, so personalized results still cover multiple interpretations of an ambiguous query.
  • The framework yields directly testable hypotheses H1–H4, so a reader could implement the architecture and check, for example, whether independent specialists with k≤2 really beat debate on simple queries, or whether bandit routing adapts to task drift within 3–5 interactions.

Where Pith is reading between the lines

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

  • The paper leaves implicit that the same persona-routing machinery could generalize beyond search, for example to task-oriented assistants that route between tool-using agents, because the persona formalism does not depend on the underlying retrieval source; testing this would require applying the coordinator to a non-search agent benchmark.
  • Because the paper reports no experimental results, its strongest claims remain open; a natural first experiment is an offline comparison of bandit routing versus static routing on a session log, measuring session utility and query-reformulation reduction.
  • A notable tension the authors acknowledge but do not resolve: persistent semantic memory enables personalization and also creates documented vulnerabilities such as memory-extraction attacks. A concrete extension would measure the privacy–utility tradeoff of their proposed mitigations, e.g., whether storing abstracted embeddings instead of raw text degrades personalization quality.
  • The coordination-protocol tradeoff could be formalized as a cost-quality frontier; an empirical mapping from query-complexity features to the optimal protocol (independent, relay, or debate) would convert the paper's qualitative predictions into a deployable routing policy.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. SPARK proposes a multi-agent framework for personalized web search. It formalizes a persona space (role, expertise, task context, domain), introduces a Persona Coordinator that routes queries to specialized LLM agents via a softmax policy, and defines coordination protocols (independent specialists, relay, constrained debate) alongside a tripartite memory subsystem and an arbiter that fuses ranked results. The paper claims that this architecture yields emergent personalization from minimal coordination rules, and it outlines four hypotheses (H1–H4) together with a planned evaluation using TREC Session Track, MS MARCO, offline counterfactual estimation, and live-user studies. No experimental results are reported; Section 5 is entirely an evaluation plan.

Significance. If validated, the framework would integrate retrieval-augmented generation, multi-agent coordination, and cognitive memory in a way that is currently missing from personalized search literature. The paper's strengths are its architectural clarity, the explicit separation of working/episodic/semantic memory, the concrete coordination protocols, and a thoughtful treatment of privacy and bias risks in Sections 6 and 7. The hypotheses are stated in falsifiable form, and the proposed ablations (e.g., removing episodic memory, disabling debate, varying k) are sensible. However, the central claim of 'emergent personalization' is asserted rather than demonstrated; the formal content is limited to a softmax routing equation and pseudocode, and the evaluation section is a plan with no results, baselines, or error bars. The load-bearing assumption that a contextual-bandit coordinator can learn reliable routing under cold start and task drift is unexamined.

major comments (4)
  1. [§5 (Evaluation), Abstract] The claim that SPARK 'demonstrates' emergent personalization is not supported by any empirical evidence. Section 5 is entirely a plan: hypotheses, proposed datasets, and intended metrics, with no experiments, baselines, or error bars. This is load-bearing because the headline contribution is an empirical claim about emergent behavior; as written, the framework is untestable. Either the paper must report results from the described setup, or it must be explicitly reframed as an untested position/framework paper whose predictions remain open.
  2. [§4.3, §4.5, H2] The Persona Coordinator is a contextual bandit over a combinatorial action space (subsets of persona agents × protocols). The paper neither specifies the reward signal nor gives a learning or regret analysis. H2 asserts adaptation to task drift 'within three to five interactions,' but no derivation, citation, or simulation justifies that rate. Since the abstract attributes emergent personalization to the coordinator's 'minimal coordination rules,' this gap is load-bearing. Provide at least a stylized cold-start analysis or a simulation before asserting adaptation.
  3. [§3 (Formal Problem)] The formal content reduces to the softmax routing equation and RRF fusion. There is no formal definition of what 'emergent personalization' means, nor of the conditions (agent diversity, memory persistence, coordination protocol) under which it is predicted to arise. The statement that SPARK 'models how emergent personalization properties arise' is hence unsupported by the formal development. A precise statement of the claimed emergence—ideally with a toy model or counterexample—is needed.
  4. [§4.5 (Adaptive Routing)] The claim that LinUCB/Thompson sampling 'has been shown in large-scale personalization systems' ([7,26]) is a mismatch: those references concern news recommendation, not multi-agent protocol selection. Also, the 'delayed reward signal' is not formally modeled, nor is exploration bounded under the budget B in Algorithm 2. Specify the learning objective, the action representation, and a stylized model of delayed feedback before claiming that the coordinator can balance exploration and exploitation.
minor comments (5)
  1. [§3] The notation ψ(π)∈R^k introduces an undefined embedding dimension k; the later use of 'top-k' is related but not explicitly tied to this k.
  2. [References] References [14] and [15] are the same paper (Dou, Song, Wen 2007), listed twice.
  3. [References] Reference [34] contains 'pp–83–95' with an unusual dash/format; please correct.
  4. [§4.4 / Algorithm 2] The free parameters θ and r_max (debate rounds) are never given default values or a discussion of sensitivity; if they are intended to be tuned, say so explicitly.
  5. [Figure 1 caption] The example 'finding LLM evaluation methods across domains' is domain-specific; consider generalizing or clarifying that it illustrates one use case.

Circularity Check

0 steps flagged

No circular derivation: SPARK is an architecture proposal with hypotheses, not fitted predictions; the only self-citations are background and non-load-bearing.

full rationale

SPARK is a framework/position paper rather than an empirical derivation. The formal model in Section 3 defines the persona tuple and the coordinator routing rule w_t(pi)=softmax(W phi(q,c)·psi(pi)) as a proposed mechanism; no parameter is fitted to data that is later relabeled as a prediction. Hypotheses H1-H4 in Section 5 are ex-ante expectations, not outputs of the model, so they cannot be equivalent to their inputs by construction. The abstract's 'testable predictions' are not obtained by fitting or by self-referential definition. The only author self-citations, [12] (Das 2023) and [34] (Oliaee/Das/Le 2025), attach to generic background sentences such as 'Recent advances in large language models (LLMs) motivate a rethinking of personalization...' and 'emergent personalization behavior...'. These references are not load-bearing: removing them would not change the formal architecture or any hypothesis. There is no uniqueness theorem, no imported ansatz, and no renaming of a known empirical pattern as if it were a derivation. Section 7's limitations concerning coordination overhead, drift, privacy, and evaluation are evidence/correctness concerns, not circularity. No specific reduction can be exhibited, so the low score reflects only minor non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

No numerical claims are made, so free parameters are architectural hyperparameters. The framework rests on several unvalidated assumptions about LLM agent behavior, debate benefits, and bandit learning. No new physical entities are introduced.

free parameters (4)
  • k (number of active personas)
    Top-k agent activation threshold; no value or fitting procedure specified in the paper.
  • r_max (debate rounds)
    Budget for bounded debate; no value specified.
  • theta (gate confidence threshold)
    Threshold for falling back from debate to independent; no value specified.
  • RRF offset = not specified
    Tunable constant in Reciprocal Rank Fusion; standard value is 60, but paper does not fix it.
axioms (5)
  • domain assumption LLM agents can reliably follow role instructions and perform retrieval-augmented reasoning
    Entire agent architecture in Section 4 assumes LLM agents can act as distinct personas without role confusion or context loss.
  • domain assumption Multi-agent debate improves factual accuracy and reasoning (Du et al. 2023; Liang et al. 2023)
    Section 2.2 and Section 4.4 adopt debate improvements as a premise for the constrained-debate protocol.
  • domain assumption Separate memory stores (working/episodic/semantic) improve retrieval efficiency and personalization fidelity
    Sections 2.4 and 4.1 motivate tripartite memory from cognitive architectures; no in-system validation.
  • domain assumption Contextual bandits can learn routing policies from implicit feedback (clicks, dwell)
    Sections 4.3 and 4.5 rely on bandit learning to select personas/protocols; feasibility is cited from prior work, not demonstrated.
  • domain assumption Reciprocal Rank Fusion is stable under heterogeneous agent scores
    Section 4.6 defaults to RRF based on Cormack et al., but its performance in this multi-agent setup is untested.
invented entities (2)
  • Persona Coordinator no independent evidence
    purpose: Routes queries to persona agents and selects collaboration protocol (independent/relay/debate)
    A software component, not a physical entity; no falsifiable handle beyond the system itself.
  • Persona space P = (role, expertise, task context, domain) no independent evidence
    purpose: Formal representation of agent specializations
    An abstract formalism; no independent empirical anchor.

pith-pipeline@v1.3.0-alltime-deepseek · 13236 in / 11689 out tokens · 101110 ms · 2026-08-03T13:27:20.256560+00:00 · methodology

0 comments
read the original abstract

Personalized search demands the ability to model users' evolving, multi-dimensional information needs; a challenge for systems constrained by static profiles or monolithic retrieval pipelines. We present SPARK (Search Personalization via Agent-Driven Retrieval and Knowledge-sharing), a framework in which coordinated persona-based large language model (LLM) agents deliver task-specific retrieval and emergent personalization. SPARK formalizes a persona space defined by role, expertise, task context, and domain, and introduces a Persona Coordinator that dynamically interprets incoming queries to activate the most relevant specialized agents. Each agent executes an independent retrieval-augmented generation process, supported by dedicated long- and short-term memory stores and context-aware reasoning modules. Inter-agent collaboration is facilitated through structured communication protocols, including shared memory repositories, iterative debate, and relay-style knowledge transfer. Drawing on principles from cognitive architectures, multi-agent coordination theory, and information retrieval, SPARK models how emergent personalization properties arise from distributed agent behaviors governed by minimal coordination rules. The framework yields testable predictions regarding coordination efficiency, personalization quality, and cognitive load distribution, while incorporating adaptive learning mechanisms for continuous persona refinement. By integrating fine-grained agent specialization with cooperative retrieval, SPARK provides insights for next-generation search systems capable of capturing the complexity, fluidity, and context sensitivity of human information-seeking behavior.

Figures

Figures reproduced from arXiv: 2512.24008 by Gaurab Chhetri, Subasish Das, Tausif Islam Chowdhury.

Figure 1
Figure 1. Figure 1: Traditional manual search is slow, fragmented, and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 14 linked inside Pith

  1. [1]

    Shipra Agrawal and Navin Goyal. 2013. Thompson sampling for contextual bandits with linear payoffs.Proceedings of the 30th International Conference on Machine Learning (ICML)(2013), 127–135

    2013

  2. [2]

    Anderson, Daniel Bothell, Michael D

    John R. Anderson, Daniel Bothell, Michael D. Byrne, Scott Douglass, Christian Lebiere, and Yulin Qin. 2004. An integrated theory of the mind.Psychological Review111, 4 (2004), 1036–1060

    2004

  3. [3]

    Jaime Arguello, Jamie Callan, and Fernando Diaz. 2011. Federated search of structured data. InProceedings of the 34th ACM SIGIR Conference on Research and Development in Information Retrieval. 83–92

    2011

  4. [4]

    Alan Baddeley. 2000. The episodic buffer: a new component of working memory? Trends in Cognitive Sciences4, 11 (2000), 417–423

    2000

  5. [5]

    2022.Neural Approaches to Personalized Search

    Emanuele Bassani. 2022.Neural Approaches to Personalized Search. Ph. D. Disser- tation. University of Milano-Bicocca

    2022

  6. [6]

    Nicholas Carlini et al. 2023. Extracting Training Data from Diffusion Models. arXiv preprint arXiv:2301.13188(2023)

    Pith/arXiv arXiv 2023

  7. [7]

    Olivier Chapelle and Lihong Li. 2011. An empirical evaluation of Thompson sampling.Advances in Neural Information Processing Systems (NeurIPS)(2011), 2249–2257

    2011

  8. [8]

    Olivier Chapelle, Donald Metzler, Ya Zhang, and Pepe Grinspan. 2011. Intent- Based Diversification of Web Search Results.Information Retrieval14, 6 (2011), 572–592

    2011

  9. [9]

    Olivier Chapelle and Ya Zhang. 2011. Intent-aware expected reciprocal rank for diversified search. InProceedings of the 34th International ACM SIGIR Conference on Research and Development in Information Retrieval. 593–602

    2011

  10. [10]

    Charles L. A. Clarke, M. Kolla, and Olga Vechtomova. 2008. An iterative approach to relevance feedback. InProceedings of the 31st ACM SIGIR Conference on Research and Development in Information Retrieval. 19–26

    2008

  11. [11]

    Cormack, Charles L

    Gordon V. Cormack, Charles L. A. Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd ACM SIGIR Conference on Research and Development in Information Retrieval. 758–759

    2009

  12. [12]

    2023.Artificial Intelligence in Highway Safety

    Subasish Das. 2023.Artificial Intelligence in Highway Safety. CRC Press, Boca Raton, FL

    2023

  13. [13]

    C. Dharap. 1998. Context-Based Profile Personalization. InProceedings of the AAAI 1998 Workshop on Personalization

    1998

  14. [14]

    Zhicheng Dou, Ruihua Song, and Ji-Rong Wen. 2007. A large-scale evaluation and analysis of personalized search strategies. InProceedings of the 16th International World Wide Web Conference (WWW). ACM, 581–590

    2007

  15. [15]

    Zhicheng Dou, Ruihua Song, and Ji-Rong Wen. 2007. A Large-Scale Evalua- tion and Analysis of Personalized Search Strategies. InProceedings of the 16th International Conference on World Wide Web

    2007

  16. [16]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch

  17. [17]

    Miroslav Dudík, John Langford, and Lihong Li. 2011. Doubly robust policy evaluation and learning. InProceedings of the 28th International Conference on Machine Learning (ICML). 1097–1104

    2011

  18. [18]

    Bennett, and Susan Dumais

    Carsten Eickhoff, Kevyn Collins-Thompson, Paul N. Bennett, and Susan Dumais

  19. [19]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. Calibrated probabilistic modeling for deep learning. InProceedings of the 34th International Conference on Machine Learning (ICML). 1321–1330

    2017

  20. [20]

    Yangyang Guo, Zhiyong Cheng, Liqiang Nie, et al. 2018. Attentive Long Short- Term Preference Modeling for Personalized Product Search.arXiv preprint arXiv:1811.10155(2018)

    Pith/arXiv arXiv 2018

  21. [21]

    Chen Jiang, Xuechen Pan, Guanzhi Hong, Chao Bao, and Min Yang. 2024. RAG- Thief: Scalable Extraction of Private Data from Retrieval-Augmented Generation Applications with Agent-Based Attacks.arXiv preprint arXiv:2411.14110

    Pith/arXiv arXiv 2024

  22. [22]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open- domain question answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 6769–6781

    2020

  23. [23]

    uttler, Mike Lewis, Wen-tau Yih, Tim Rockt

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K"uttler, Mike Lewis, Wen-tau Yih, Tim Rockt"aschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 33. 9459–9474

    2020

  24. [24]

    Guohao Li, Hasan A. A. K. Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. Camel: Communicative agents for “mind” exploration of large lan- guage model society. InProceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS)

    2023

  25. [25]

    Haotian Li, Yifan Xu, Sherry Guo, and Vincent Tan. 2024. SafeAgentBench: Benchmarking Safety, Robustness, and Coordination in Multi-Agent LLM Sys- tems.arXiv preprint arXiv:2409.06789

    Pith/arXiv arXiv 2024

  26. [26]

    Schapire

    Lihong Li, Wei Chu, John Langford, and Robert E. Schapire. 2010. A contextual- bandit approach to personalized news article recommendation. InProceedings of the 19th International Conference on World Wide Web (WWW). 661–670

    2010

  27. [27]

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Zhaopeng Tu, and Shuming Shi. 2023. Encouraging divergent thinking in large language models through multi-agent debate.arXiv preprint arXiv:2305.19118(2023)

    Pith/arXiv arXiv 2023

  28. [28]

    David Martyr. 2025. Complexity Does Not Always Equal Competence: Coor- dination Overhead in Multi-Agent LLM Systems. InProceedings of the AAAI Conference on Artificial Intelligence

    2025

  29. [29]

    Anya Mirchandani, Brandon Lee, and Philip Wong. 2024. DAWN: Dynamic Agent Workflow Navigation for Adaptive Reasoning.arXiv preprint arXiv:2408.01234

    Pith/arXiv arXiv 2024

  30. [30]

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, et al. 2021. WebGPT: Browser-assisted question-answering with human feedback. InAdvances in Neural Information Processing Systems

    2021

  31. [31]

    Karthik Narasimhan, Adam Yala, and Regina Barzilay. 2016. Learning to search in long documents using document structure. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. 1203–1212

    2016

  32. [32]

    Tien T Nguyen, Pik-Mai Hui, F Maxwell Harper, Loren Terveen, and Joseph A Konstan. 2014. Exploring the Filter Bubble: The Effect of Using Recommender Systems on Content Diversity. InProceedings of the 23rd International World Wide Web Conference. 677–686

    2014

  33. [33]

    Rodrigo Nogueira and Kyunghyun Cho. 2019. Multi-stage document ranking with BERT. InarXiv preprint arXiv:1910.14424

    Pith/arXiv arXiv 2019

  34. [34]

    Amir Hossein Oliaee, Subasish Das, and Minh Le. 2025. Automating Pedestrian Crash Typology Using Transformer Models.Transportation Research Record: Journal of the Transportation Research Board2679, 2 (Feb. 2025), pp–83–95. doi:10. 1177/03611981241260691

    2025

  35. [35]

    Santiago Ontañón and Jichen Zhu. 2021. The Personalization Paradox: the Conflict between Accurate User Models and Personalized Adaptive Systems. arXiv preprint arXiv:2103.01771(2021)

    Pith/arXiv arXiv 2021

  36. [36]

    2011.The Filter Bubble: What the Internet is Hiding from You

    Eli Pariser. 2011.The Filter Bubble: What the Internet is Hiding from You. Penguin UK. SPARK: Search Personalization via Agent-Driven Retrieval and Knowledge-sharing WSDM Companion ’26, February 22–26, 2026, Boise, ID, USA

    2011

  37. [37]

    Baolin Peng, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Model Editing and Unlearning in Large Language Models: Methods, Challenges, and Opportunities.arXiv preprint arXiv:2307.05785(2023)

    Pith/arXiv arXiv 2023

  38. [38]

    Erasmo Purificato, Ludovico Boratto, and Ernesto William De Luca. 2024. User Modeling and User Profiling: A Comprehensive Survey.arXiv preprint arXiv:2402.09660(2024)

    Pith/arXiv arXiv 2024

  39. [39]

    Matthijs Radlinski and Thorsten Joachims. 2011. Personalizing Web Search using Long Term Browsing History. InWSDM

    2011

  40. [40]

    Daniel Russo, Benjamin Van Roy, Abbas Kazerouni, Ian Osband, and Zheng Wen

  41. [41]

    Milad Sabouri et al . 2025. Temporal user profiling with LLMs: Balancing short-term and long-term preferences for recommendations.arXiv preprint arXiv:2508.08454(2025)

    Pith/arXiv arXiv 2025

  42. [42]

    Christian Sandvig, Kevin Hamilton, Karrie Karahalios, and Cedric Langbort

  43. [43]

    Weijia Shi and et al. 2023. Prompt Injection Attacks against Large Language Models.arXiv preprint arXiv:2303.08987(2023)

    Pith/arXiv arXiv 2023

  44. [44]

    Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. InFindings of the Association for Computational Linguistics: EMNLP 2021. 3784–3803

    2021

  45. [45]

    John Smith, Alice Lee, and Rahul Gupta. 2022. Fairness in Personalization: Mitigating Bias and Promoting Diversity in Recommender Systems.Comput. Surveys(2022)

    2022

  46. [46]

    John Smith, Aisha Patel, and Kevin Wong. 2021. Multi-agent coordination for adaptive retrieval strategies. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 1234– 1245

    2021

  47. [47]

    Ben Steichen and Bo Fu. 2019. Towards Adaptive Information Visualization: A Study of Visualization Aids and the Role of User Cognitive Style.ACM Transac- tions on Computer-Human Interaction (TOCHI)(2019)

    2019

  48. [48]

    Adith Swaminathan and Thorsten Joachims. 2015. Counterfactual risk min- imization: Learning from logged bandit feedback. InProceedings of the 32nd International Conference on Machine Learning (ICML). 814–823

    2015

  49. [49]

    Dumais, and Eric Horvitz

    Jaime Teevan, Susan T. Dumais, and Eric Horvitz. 2005. Personalizing search via automated analysis of interests and activities. InProceedings of the 28th ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 449–456

    2005

  50. [50]

    Bo Wang, Weiyi He, Shenglai Zeng, Zhen Xiang, Yue Xing, Jiliang Tang, and Pengfei He. 2025. Unveiling Privacy Risks in LLM Agent Memory. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics

    2025

  51. [51]

    Xiao Wang, Yuxuan Li, Jie Zhang, et al. 2023. Plan-and-Solve: Prompting large language models to perform multi-step reasoning. InProceedings of ACL

    2023

  52. [52]

    Z Yu et al. 2019. Adaptive user modeling with long and short-term preferences. InIJCAI

    2019

  53. [53]

    Ming Yuan, Fei Zhao, and Han Liu. 2024. Privacy-Preserving Retrieval-Augmented Generation.arXiv preprint arXiv:2402.12345

    Pith/arXiv arXiv 2024

  54. [54]

    Jijie Zhou, Eryue Xu, Yaoyao Wu, and Tianshi Li. 2025. Rescriber: Smaller-LLM- Powered User-Led Data Minimization for LLM-Based Chatbots. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Received 27 November 2025

    2025

  55. [2013]

    InProceedings of the 6th ACM International Conference on Web Search and Data Mining (WSDM)

    Personalizing atypical web search sessions. InProceedings of the 6th ACM International Conference on Web Search and Data Mining (WSDM). ACM, 285–294

  56. [2014]

    InProceedings of the 64th Annual Meeting of the International Communication Association

    Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms. InProceedings of the 64th Annual Meeting of the International Communication Association

  57. [2018]

    A tutorial on Thompson sampling.Foundations and Trends in Machine Learning11, 1 (2018), 1–96

    2018

  58. [2023]

    Improving factuality and reasoning in language models through multi-agent debate.arXiv preprint arXiv:2305.14325(2023)

    Pith/arXiv arXiv 2023