Pith. sign in

REVIEW 4 major objections 5 minor 59 references

Benchmarking the Personalization Capabilities of Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read Current LLMs recover at most about half of the strategic content humans use to close sales deals.

desk verdict A solid new two-party benchmark for sales personalization, with a plateau claim that currently rests on missing statistics; the email-cohort null needs tests before publication. read the letter →

arxiv 2607.20471 v1 pith:AI7L23AS submitted 2026-05-23 cs.AI

classification cs.AI
keywords personalizationBayesianpersuasionlargelanguagemodelssalesoutreachbenchmarkingcustomersuccessstoriespitchpointstemporalleakage
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

This paper tries to establish whether large language models can perform personalization in the classic two-party sense: varying a message to induce a specific receiver to act. Using sales outreach as the testbed, the authors adapt Bayesian Persuasion to generative agents and build SDR-Bench, a corpus of 6,279 customer success stories plus real emails and call transcripts with downstream outcomes. They find a consistent personalization plateau: the best model recovers about 56% of the strategic pitch points in human deal-closing stories, and on a Fortune 100 tech email cohort no model statistically ranks successful outreach above unsuccessful outreach. A field deployment with 12 sales representatives rated 48% of model-generated pitch points immediately usable. If the plateau holds, it means autonomous LLM personalization for high-stakes persuasion is not yet viable, and the paper provides a reproducible way to measure progress toward it.

What carries the argument

The central device is Bayesian Persuasion recast as a signaling game between a sender (the LLM agent) and a receiver (the prospect), with unobserved receiver state decomposed into needs and wants. Since the receiver's utility is unobservable, the authors substitute an observable proxy: the Weighted Coverage Score, the semantic overlap between an agent's pitch points and the triadic pitch points (product → pain point → value mechanism) extracted from a human message known to have induced action. Ground-truth extraction uses an LLM judge with 0.92 precision and 0.97 recall against human annotators. A historical-internet simulator restricts agents to information published before the interaction

What would settle it

A randomized field experiment: generate two sets of outreach for the same prospects, one scoring high and one scoring low on WCS, and log actual reply/call/meeting rates. If high-WCS messages do not induce higher action rates than low-WCS messages, the claim that WCS tracks the Bayesian persuasion objective is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that current LLMs generate fluent personalized messages but do not yet solve two-party personalization as formalized by Bayesian Persuasion. The authors treat a human-authored outreach that triggered a known receiver action as a sample of a utility-maximizing message, extract its pitch points, and define the Weighted Coverage Score (WCS) as the fraction an agent's output semantically covers. Across frontier LLMs and deep-research agents, WCS values cluster in the 30–43% range on aggregate evaluation sets, with the strongest model reaching 55.8% on public success stories, and no model separates successful from unsuccessful emails in the technology cohort. The auth

Load-bearing premise

The whole evaluation assumes that pitch points extracted from a human success story are a faithful and complete record of what persuaded the receiver, so semantic overlap with those points measures persuasive quality; if that content is not causally linked to the observed action, WCS scores stop meaning what the paper claims.

Editorial extensions

If this is right

  • If the plateau is real, autonomous LLM sales outreach in competitive high-stakes sectors is not ready; models can produce coherent but strategically shallow messages.
  • Deep-research agents reach the upper end of the WCS range but at one to two orders of magnitude greater inference cost, so frontier LLMs with search are the more compute-efficient frontier.
  • WCS provides a reproducible benchmark for comparing generative personalization systems without rerunning expensive human A/B studies on each new model.
  • The negligible pre/post-2024 WCS difference implies the plateau reflects context-conditioned synthesis rather than memorized success stories.

Reading between the lines

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

  • If the WCS proxy is accepted, the plateau likely extends beyond sales to any domain where persuasion depends on matching a receiver's decision-relevant needs, but that extension is my inference, not the paper's.
  • Because WCS rewards overlap with one successful story, agents that find equally valid but different winning arguments will be under-scored; a paired benchmark against unsuccessful stories as a lower bound could reveal this.
  • The most direct test not run in the paper is a randomized field experiment comparing high-WCS versus low-WCS model messages on actual receiver action rates; that would connect the proxy to the persuasion objective it stands in for.
  • A future benchmark could use the same temporal simulation to test whether fine-tuning on WCS improves downstream action rates, turning the measure into a training signal rather than only an evaluation metric.
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

4 major / 5 minor

Summary. The paper introduces SDR-Arena, a benchmark for evaluating LLM performance on two-party generative personalization, instantiated in sales outreach. The authors cast personalization as a Bayesian Persuasion task and define the Weighted Coverage Score (WCS), which measures semantic overlap between an agent's generated pitch points and pitch points extracted from human-authored artifacts known to have induced a receiver action. They release SDR-Bench, a public corpus of 6,279 customer success stories, plus a private enterprise email/transcript corpus, and evaluate several LLMs and deep-research agents under a temporally constrained web-search simulation. Reported results show a 'personalization plateau' with WCS scores concentrated around 30–56%, and a finding that on the Fortune 100 tech cohort no model statistically separates successful from unsuccessful outreach. The paper also reports human validation studies: extraction F1=0.95, judge–human Spearman ρ=0.74, per-pitch usefulness 48%, and senior-SDR overlap Pearson r=0.82.

Significance. If the central claims are sustained, this would be a useful and unusual contribution: a reproducible, temporally controlled benchmark for a downstream behavioral objective (inducing receiver action), rather than a purely text-similarity benchmark. The public release of SDR-Bench and the SDR-Arena framework, the historical web simulator to prevent future leakage, and the field deployment with professional SDRs are genuine strengths and are appropriately credited. The human-validation results provide partial independent ground truth for both extraction and judging. However, the headline negative result—that no model statistically separates successful from unsuccessful tech-cohort outreach—is currently unsupported by the reported statistics, and the proxy-validity assumption linking WCS to receiver action is asserted rather than tested. These are load-bearing concerns because the plateau and the benchmark's practical meaning rest on them.

major comments (4)
  1. [§4.1, Table 2, Appendix A.10] The claim that 'no model statistically separates successful from unsuccessful outreach' in the Technology cohort is not supported by any reported significance test, confidence interval, or effect-size statistic. Table 2 reports point estimates only (e.g., STORM 43.15 vs. 39.24; GPT-5.4 53.02 vs. 53.80). With roughly 200 emails per arm, a 3–5 WCS-point difference can easily arise from sampling noise. Appendix A.10 provides bootstrap CIs only for SDR-Bench and only for STORM, ODR, GPT-4o, and Qwen; it does not cover Claude Sonnet 4.6, GPT-5.4, the enterprise email cohorts, or the successful/unsuccessful comparison. The central negative finding therefore needs either pairwise significance tests with appropriate multiple-comparison control, or explicit equivalence tests, on the email-cohort data. This is directly fixable with data already in the authors' possession.
  2. [§2.3, §3.1] The empirical proxy equates WCS—semantic overlap between agent output and pitch points extracted from a human message that induced action—with the Bayesian Persuasion objective of inducing receiver action. This is an assumption, not a derived equivalence. The validation studies show that WCS correlates with expert judgment (Pearson 0.82) and that a judge tracks human ratings (Spearman 0.74), but they do not show that high WCS causes, or even predicts, receiver action. Since the paper itself has success/failure labels in the email cohorts, a concrete and feasible test would be to check whether human-authored successful emails achieve reliably higher WCS than unsuccessful emails when scored against a held-out successful story or another success-labeled reference, and whether an agent's WCS differential across cohorts tracks the human differential. Without such a test, the interpretability
  3. [§3.1, Appendix A.7 (Pitch Point Generation Prompt)] The agent prompt instructs the model to 'generate 3 targeted pitch points,' while ground-truth pitch points are extracted without such a cap and can be more numerous. Because WCS is computed as the fraction of ground-truth points covered, the fixed output length imposes an upper bound on achievable WCS. If the average number of ground-truth pitch points substantially exceeds 3, this alone could compress scores into the observed 30–56% range and contribute to the 'plateau.' The paper should report the distribution of |V*|, and either vary the requested number of pitch points, or normalize by precision/recall in a way that is robust to output length. As written, the plateau may partly reflect an artifact of the prompting protocol rather than a property of the models.
  4. [Appendix A.10] The bootstrap analysis is used to 'validate the personalization plateau,' but overlapping confidence intervals for GPT-4o and Qwen do not establish that the models are equivalent; they only show the absence of a significant difference under a limited sample. To support a plateau claim, the paper would need either a formal equivalence test with pre-specified bounds, or a demonstration that adding more capable models does not move the scores. Furthermore, the CIs in Table 6 are computed on a subset of the SDR-Bench evaluation set and exclude several evaluated models (Claude Sonnet 4.6, GPT-5.4, GPT-5.4-mini, GPT-4o-mini). This is an incomplete basis for the paper's central negative conclusion.
minor comments (5)
  1. [References] Terho et al. (2022a) and (2022b) appear to be the same publication; one reference should be removed or the two distinct works should be disambiguated.
  2. [Table 1] The last row, 'Golden dataset handpicked400 400,' is typographically unclear; presumably it means 400 per class for each enterprise. Please clarify the formatting.
  3. [Appendix A.2.1] Typo: 'comparision' should be 'comparison.'
  4. [Appendix A.2.2] 'M&A' is written as 'MA' in the Event-based strategy description; please correct.
  5. [§3.1] The WCS formula is written as 'WCS = (Σ s_i / 5N) × 100%'; it may be clearer to define N as the number of ground-truth pitch points and specify the handling of missing candidate matches, since the judge can output a score of 0 for a missed point.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the WCS proxy is an explicitly acknowledged approximation and is externally validated against human judgments.

full rationale

The paper's derivation chain is: adapt Bayesian Persuasion to generative agents; replace the unobservable receiver utility with an observable proxy (WCS), defined as semantic overlap between agent pitch points and pitch points extracted from human-authored artifacts known to have induced action. WCS is not defined in terms of the evaluated models' outputs, and no parameter is fitted to a subset and then predicted as a 'prediction.' Ground-truth pitch points are extracted from human artifacts, not from model outputs, and the coverage judge scores generated pitch points against those human-derived points. Although GPT-4o serves both as the extractor/judge and is itself evaluated, the paper supplies independent human grounding: extraction precision/recall/F1 of 0.92/0.97/0.95 on 30 stories; judge-human rank correlation Spearman 0.7435 on 20 stories; gold-standard expert overlap with WCS at Pearson 0.816; and a field deployment in which 48.2% of model-generated pitch points were rated immediately useful by senior SDRs. These external validations break any self-referential loop. The abstract claim that 'no model statistically separates successful from unsuccessful outreach' in the tech cohort is not supported by any reported significance test in the paper, but that is a statistical-inference gap, not a circularity: it does not reduce the conclusion to the benchmark's inputs. No load-bearing self-citations or ansatz-smuggling chains are present. The central proxy assumption (overlap with successful-story content approximates persuasion) is asserted rather than derived, but it is a stated modeling choice, not a definitional identity, and the field validation provides independent evidence for it.

Assumptions & free parameters 1 free parameters · 4 assumptions · 2 invented entities

The benchmark's numbers depend on a chain of assumptions: the rational-receiver Bayesian model, the identification of success-story content with the set of utility-maximizing messages, the reliability of an LLM judge that is itself an LLM, and the temporal integrity of the search simulation. The only fitted/hand-chosen quantity affecting the main scores is the fixed output length (3 pitch points), which structurally caps WCS.

free parameters (1)
  • number of pitch points requested per agent = 3 (prompt default)
    The Pitch Point Generation Prompt asks for '3 targeted pitch points'; this caps WCS at 3*5/(N*5). If ground truth has more than 3 pitch points, scores cannot exceed 3/N, so the 'plateau' may reflect output-length constraints rather than model capability. No sensitivity analysis is reported.
assumptions (4)
  • domain assumption Bayesian Persuasion model of receiver: receiver is rational Bayesian with prior mu and reservation threshold tau, acting only when expected utility exceeds tau
    Invoked in Section 2.2; standard economics but assumed to hold for B2B prospects, with no empirical test in this paper.
  • domain assumption Successful human artifacts are samples from utility-maximizing messages; content overlap (WCS) is a lower-bound proxy for the persuasion objective
    Section 2.3 states the proxy replacement; this is the load-bearing assumption that grounds all WCS scores. It is asserted, with validation only via expert agreement, not receiver actions.
  • domain assumption LLM judge faithfully measures semantic alignment on the 0-5 rubric
    Used throughout; validated against humans on 20 stories (Spearman 0.74), but the judge is systematically more conservative, and email-cohort scoring has no human validation.
  • domain assumption SERP API date parameters enforce a strict historical web boundary
    Section 3.1 'Historical Internet Simulator'; assumes search-engine date filtering excludes post-t information reliably, which is not demonstrated.
invented entities (2)
  • Pitch point triad structure (Product/Service -> Pain Point -> Value Proposition/Mechanism) independent evidence
    purpose: Decomposes messages into comparable atomic arguments so coverage can be scored
    Validated against expert annotations on 30 stories (precision 0.92, recall 0.97), so it has a falsifiable handle outside the paper.
  • omega_t = {needs n_i, wants w_i} latent receiver state
    purpose: Formalizes the unobserved compatibility between receiver requirements and seller product
    Introduced in Section 2.2 as a theoretical construct; never directly measured or estimated, so it has no independent falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking the Personalization Capabilities of Large Language Models." pith.science (2026). https://pith.science/paper/AI7L23AS

@misc{pith2026260720471,
  author       = {Pith},
  title        = {Pith review of: Benchmarking the Personalization Capabilities of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AI7L23AS}},
  note         = {Machine review of arXiv:2607.20471}
}
read the original abstract

Personalization, the act of varying a message to induce action from a specific receiver while keeping sender, channel, and time fixed, has a long tradition in psychology and marketing as a two-party problem in which sender and receiver have independent objectives. Large language models remove the bounded-inventory constraint of classical retrieval-and-ranking approaches by generating a continuum of message variants conditioned on inferred receiver state, raising the question of how well current models perform personalization in the classical sense. Existing LLM personalization benchmarks measure sender-side adaptation, in which the receiver is the same user the model is serving. The two-party question, whether a generated message induces its intended action in a third party, has been investigated only through A/B tests and small-scale human studies that cannot be re-run against a new model on demand. We adapt the Bayesian Persuasion framework of Kamenica and Gentzkow (2011) to generative agents and instantiate the formulation in sales, where receiver actions are routinely logged against the outreach that induced them. We release SDR-Bench, a public corpus of 6,279 customer success stories spanning 22 industries and approximately 200 enterprises, served through a temporally constrained simulation that prevents future-data leakage. Across frontier LLMs and deep-research agents, we observe a consistent personalization plateau and on a Fortune 100 tech cohort no model statistically separates successful from unsuccessful outreach. A field deployment with 12 professional sales representatives validates the framework, with 48 percent of model-generated content rated immediately useful and senior-expert agreement at Pearson 0.82. We release SDR-Arena and SDR-Bench publicly to support reproducible study of generative personalization at scale.

Figures

Figures reproduced from arXiv: 2607.20471 by the authors.

Figure 1
Figure 1. Overview of SDR-Arena showcasing how LLM generated output is compared with artifacts [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Sales Journey: From Prospecting to Outreach to Call and eventual Deal Closure leading to [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution of count of strategies across a random subset of 34,000 emails [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Personalization in actual Sales Emails [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Distribution of count of Success Stories by Industry Type [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Qualitative Example: Ground truth pitch points scored against pitch points generated by the agent A.1 SDR-Bench: Dataset Curation Details Filtration Criteria Count Domains Found for Companies with over $1B revenue ∼30k Domains Found for B2B Companies with over $1B reve…
Figure 8
Figure 8. Figure 8: Graph of token usage vs performance of various agents [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 1 canonical work pages

  1. [1]

    2025 , eprint=

    DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents , author=. 2025 , eprint=

  2. [2]

    2011 , publisher=

    Predictable Revenue: Turn Your Business Into a Sales Machine with the \ 100 Million Best Practices of Salesforce.com , author=. 2011 , publisher=

  3. [3]

    1953 , publisher=

    Communication and Persuasion: Psychological Studies of Opinion Change , author=. 1953 , publisher=

  4. [4]

    1990 , organization =

    Napoli, Jack and Dunkel, Richard , title =. 1990 , organization =

  5. [5]

    Toward a contextualized understanding of inside sales: the role of sales development in effective lead funnel management , volume =

    Terho, Harri and Salonen, Anna and Yrjänen, Meri , year =. Toward a contextualized understanding of inside sales: the role of sales development in effective lead funnel management , volume =. Journal of Business and Industrial Marketing , doi =

  6. [6]

    The communication of ideas , volume=

    The structure and function of communication in society , author=. The communication of ideas , volume=. 1948 , publisher=

  7. [8]

    Annual Review of Psychology , volume=

    Selective attention , author=. Annual Review of Psychology , volume=. 1986 , publisher=

  8. [9]

    , address =

    Oxford University Press. , address =. Oxford English Dictionary. , year =

Show all 59 references
  1. [10]

    Recommender Systems Handbook , volume =

    Ricci, Francesco and Rokach, Lior and Shapira, Bracha , year =. Recommender Systems Handbook , volume =. Recommender Systems Handbook , doi =

  2. [11]

    Journal of marketing , volume=

    Predictability and personalization in the service encounter , author=. Journal of marketing , volume=. 1987 , publisher=

  3. [12]

    Psychology & Marketing , volume=

    Personalization in personalized marketing: Trends and ways forward , author=. Psychology & Marketing , volume=. 2022 , publisher=

  4. [13]

    Online Display Advertising Markets: A Literature Review and Future Directions , volume =

    Choi, Hana and Mela, Carl and Balseiro, Santiago and Leary, Adam , year =. Online Display Advertising Markets: A Literature Review and Future Directions , volume =. Information Systems Research , doi =

  5. [14]

    Personalized Dialogue Generation with Persona-Adaptive Attention , volume =

    Huang, Qiushi and Zhang, Yu and Ko, Tom and Liu, Xubo and Wu, Bo and Tang, H , year =. Personalized Dialogue Generation with Persona-Adaptive Attention , volume =. Proceedings of the AAAI Conference on Artificial Intelligence , doi =

  6. [15]

    Recommender systems handbook , pages=

    Introduction to recommender systems handbook , author=. Recommender systems handbook , pages=. 2010 , publisher=

  7. [16]

    Artificial intelligence in marketing , pages=

    AI and personalization , author=. Artificial intelligence in marketing , pages=. 2023 , publisher=

  8. [17]

    Discover Sustainability , volume=

    The role of large language models in personalized learning: a systematic review of educational impact , author=. Discover Sustainability , volume=. 2025 , publisher=

  9. [18]

    International Journal of Artificial Intelligence in Education , pages=

    Generative AI in the classroom: effects of context-personalized learning material and tasks on motivation and performance , author=. International Journal of Artificial Intelligence in Education , pages=. 2025 , publisher=

  10. [19]

    arXiv preprint arXiv:2512.03373 , year=

    LLM-Generated Ads: From Personalization Parity to Persuasion Superiority , author=. arXiv preprint arXiv:2512.03373 , year=

  11. [20]

    World Wide Web , volume=

    When large language models meet personalization: Perspectives of challenges and opportunities , author=. World Wide Web , volume=. 2024 , publisher=

  12. [21]

    Nature Machine Intelligence , volume=

    The benefits, risks and bounds of personalizing the alignment of large language models to individuals , author=. Nature Machine Intelligence , volume=. 2024 , publisher=

  13. [22]

    American Economic Review , volume=

    Bayesian persuasion , author=. American Economic Review , volume=. 2011 , publisher=

  14. [23]

    GitHub Repository , howpublished =

    UncleCode , title =. GitHub Repository , howpublished =. 2024 , publisher =

  15. [24]

    arXiv preprint arXiv:2309.16609 , year=

    Qwen Technical Report , author=. arXiv preprint arXiv:2309.16609 , year=

  16. [25]

    Assisting in Writing W ikipedia-like Articles From Scratch with Large Language Models

    Shao, Yijia and Jiang, Yucheng and Kanell, Theodore and Xu, Peter and Khattab, Omar and Lam, Monica. Assisting in Writing W ikipedia-like Articles From Scratch with Large Language Models. Proceedings of the 2024 Conference of the North American Chapter of the Association for C...

  17. [27]

    and Peters, Heinrich and Harari, Gabriella and Cerf, M

    Matz, Sandra and Vaid, S. and Peters, Heinrich and Harari, Gabriella and Cerf, M. , year =. The potential of generative AI for personalized persuasion at scale , volume =. Scientific Reports , doi =

  18. [28]

    Recuperado el , volume=

    The future of personalization—and how to get ready for it , author=. Recuperado el , volume=

  19. [29]

    Scientific Reports , volume=

    The potential of generative AI for personalized persuasion at scale , author=. Scientific Reports , volume=. 2024 , publisher=

  20. [30]

    When large language models meet personalization: perspectives of challenges and opportunities , volume=

    Chen, Jin and Liu, Zheng and Huang, Xu and Wu, Chenwang and Liu, Qi and Jiang, Gangwei and Pu, Yuanhao and Lei, Yuxuan and Chen, Xiaolong and Wang, Xingmei and Zheng, Kai and Lian, Defu and Chen, Enhong , year=. When large language models meet personalization: perspectives of ...

  21. [31]

    Tavily AI , author =

  22. [32]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Lamp: When large language models meet personalization , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  23. [37]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Personalens: A benchmark for personalization evaluation in conversational ai assistants , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

  24. [38]

    2024 , date =

    Esin Durmus and Liane Lovitt and Alex Tamkin and Stuart Ritchie and Jack Clark and Deep Ganguli , title =. 2024 , date =

  25. [39]

    2025 , howpublished =

    LangChain , title =. 2025 , howpublished =

  26. [42]

    Bright Data SERP API , year =

  27. [43]

    Claude Sonnet 4.6 , year =

  28. [44]

    Persobench: Benchmarking personalized response generation in large language models

    Saleh Afzoon, Zahra Jamali, Usman Naseem, and Amin Beheshti. Persobench: Benchmarking personalized response generation in large language models. arXiv preprint arXiv:2410.03198, 2024

  29. [45]

    Claude sonnet 4.6

    Anthropic . Claude sonnet 4.6. https://www.anthropic.com/claude/sonnet, 2026. Accessed: 2026-05-07

  30. [46]

    Bright data serp api, 2026

    Bright Data . Bright data serp api, 2026. URL https://brightdata.com/products/serp-api. Accessed: 2026-05-07

  31. [47]

    When large language models meet personalization: Perspectives of challenges and opportunities

    Jin Chen, Zheng Liu, Xu Huang, Chenwang Wu, Qi Liu, Gangwei Jiang, Yuanhao Pu, Yuxuan Lei, Xiaolong Chen, Xingmei Wang, et al. When large language models meet personalization: Perspectives of challenges and opportunities. World Wide Web, 27 0 (4): 0 42, 2024

  32. [48]

    Online display advertising markets: A literature review and future directions

    Hana Choi, Carl Mela, Santiago Balseiro, and Adam Leary. Online display advertising markets: A literature review and future directions. Information Systems Research, 31, 06 2020. doi:10.1287/isre.2019.0902

  33. [49]

    Measuring the persuasiveness of language models, 2024

    Esin Durmus, Liane Lovitt, Alex Tamkin, Stuart Ritchie, Jack Clark, and Deep Ganguli. Measuring the persuasiveness of language models, 2024. URL https://www.anthropic.com/news/measuring-model-persuasiveness

  34. [50]

    Hovland, I.L

    C.I. Hovland, I.L. Janis, and H.H. Kelley. Communication and Persuasion: Psychological Studies of Opinion Change. Yale paperbound. Yale University Press, 1953. URL https://books.google.co.in/books?id=ZYizW6_P-goC

  35. [51]

    Personamem-v2: Towards personalized intelligence via learning implicit user personas and agentic memory

    Bowen Jiang, Yuan Yuan, Maohao Shen, Zhuoqun Hao, Zhangchen Xu, Zichen Chen, Ziyi Liu, Anvesh Rao Vijjini, Jiashu He, Hanchao Yu, et al. Personamem-v2: Towards personalized intelligence via learning implicit user personas and agentic memory. arXiv preprint arXiv:2512.06688, 2025

  36. [52]

    Bayesian persuasion

    Emir Kamenica and Matthew Gentzkow. Bayesian persuasion. American Economic Review, 101 0 (6): 0 2590--2615, 2011

  37. [53]

    Open deep research

    LangChain. Open deep research. https://github.com/langchain-ai/open_deep_research, 2025. Accessed: 2026-05-07

  38. [54]

    The structure and function of communication in society

    Harold D Lasswell. The structure and function of communication in society. In The communication of ideas, volume 37, pages 215--228. Harper and Row, 1948

  39. [55]

    A personalized conversational benchmark: Towards simulating personalized conversations

    Li Li, Peilin Cai, Ryan A Rossi, Franck Dernoncourt, Branislav Kveton, Junda Wu, Tong Yu, Linxin Song, Tiankai Yang, Yuehan Qin, et al. A personalized conversational benchmark: Towards simulating personalized conversations. arXiv preprint arXiv:2505.14106, 2025

  40. [56]

    Vaid, Heinrich Peters, Gabriella Harari, and M

    Sandra Matz, S. Vaid, Heinrich Peters, Gabriella Harari, and M. Cerf. The potential of generative ai for personalized persuasion at scale. Scientific Reports, 14, 02 2024. doi:10.1038/s41598-024-53755-0

  41. [57]

    Gpt-4 technical report

    OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. URL https://arxiv.org/abs/2303.08774

  42. [58]

    Petty and John T

    Richard E. Petty and John T. Cacioppo. The elaboration likelihood model of persuasion. volume 19 of Advances in Experimental Social Psychology, pages 123--205. Academic Press, 1986. doi:https://doi.org/10.1016/S0065-2601(08)60214-2. URL https://www.sciencedirect.com/science/ar...

  43. [59]

    Introduction to recommender systems handbook

    Francesco Ricci, Lior Rokach, and Bracha Shapira. Introduction to recommender systems handbook. In Recommender systems handbook, pages 1--35. Springer, 2010

  44. [60]

    Lamp: When large language models meet personalization

    Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. Lamp: When large language models meet personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7370--7392, 2024

  45. [61]

    Assisting in writing W ikipedia-like articles from scratch with large language models

    Yijia Shao, Yucheng Jiang, Theodore Kanell, Peter Xu, Omar Khattab, and Monica Lam. Assisting in writing W ikipedia-like articles from scratch with large language models. In Kevin Duh, Helena Gomez, and Steven Bethard, editors, Proceedings of the 2024 Conference of the North A...

  46. [62]

    The role of large language models in personalized learning: a systematic review of educational impact

    Sahil Sharma, Puneet Mittal, Mukesh Kumar, and Vivek Bhardwaj. The role of large language models in personalized learning: a systematic review of educational impact. Discover Sustainability, 6 0 (1): 0 1--24, 2025

  47. [63]

    Generative ai in the classroom: effects of context-personalized learning material and tasks on motivation and performance

    Osman Tasdelen and Daniel Bodemer. Generative ai in the classroom: effects of context-personalized learning material and tasks on motivation and performance. International Journal of Artificial Intelligence in Education, pages 1--22, 2025

  48. [65]

    Toward a contextualized understanding of inside sales: the role of sales development in effective lead funnel management

    Harri Terho, Anna Salonen, and Meri Yrjänen. Toward a contextualized understanding of inside sales: the role of sales development in effective lead funnel management. Journal of Business & Industrial Marketing, 38 0 (2): 0 337--352, 09 2022 b . ISSN 0885-8624. doi:10.1108/JBIM...

  49. [66]

    Qwen2.5 technical report

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. URL https://arxiv.org/abs/2412.15115

  50. [67]

    Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. Personalizing dialogue agents: I have a dog, do you have pets too? In Iryna Gurevych and Yusuke Miyao, editors, Proceedings of the 56th Annual Meeting of the Association for Computational L...

  51. [68]

    Personalens: A benchmark for personalization evaluation in conversational ai assistants

    Zheng Zhao, Clara Vania, Subhradeep Kayal, Naila Khan, Shay B Cohen, and Emine Yilmaz. Personalens: A benchmark for personalization evaluation in conversational ai assistants. In Findings of the Association for Computational Linguistics: ACL 2025, pages 18023--18055, 2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.