pith. sign in

arxiv: 2606.02883 · v1 · pith:4JBHW6M7new · submitted 2026-06-01 · 💻 cs.HC · cs.AI· cs.CY· cs.IR

LLM-Assisted Reranking to Operationalize Nuanced Objectives in Recommender Systems

Amir Ghasemian , Homa Hosseinmardi , Upasana Dutta , Duncan J. Watts This is my paper

Pith reviewed 2026-06-28 12:23 UTC · model grok-4.3

classification 💻 cs.HC cs.AIcs.CYcs.IR
keywords recommender systemsLLM rerankingideological diversityconspiratorial contentprompt regularizationYouTube recommendationspersonalizationzero-shot prompting
0
0 comments X

The pith

Unconstrained LLM reranking of YouTube recommendations strengthens personalization but increases exposure to conspiratorial and extremist content for users whose histories already contain it.

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

The paper examines LLM-assisted reranking of YouTube sidebar candidates drawn from real user news-consumption histories. A baseline zero-shot prompt improves how closely recommendations match past behavior yet also raises the share of conspiratorial or extreme political material shown to users who already consume such content. Adding lightweight constraints to the prompt instruction preserves topical relevance while cutting extreme content promotion, raising ideological diversity, and incurring only modest relevance loss. Synthetic tests indicate the model responds to statistical patterns in language rather than any deeper grasp of ideology. The work shows that prompt design can embed broader objectives into recommendation but must be evaluated for its unintended exposure effects.

Core claim

Without constraints, reranking strengthened personalization but increased exposure to conspiratorial and extremist material for users whose histories contained such content. Lightweight prompt-level regularization reduced promotion of extreme content and increased ideological diversity, with modest relevance loss. Synthetic experiments suggest that LLMs rerank via statistical regularities in language rather than semantic understanding of ideology.

What carries the argument

Zero-shot instruction-based prompting applied to rerank YouTube sidebar candidates, compared across an unconstrained prompt and a constrained variant that adds requirements for ideological breadth and reduced extreme content.

If this is right

  • Reranking can be used to operationalize objectives such as ideological diversity without retraining the underlying recommender.
  • Prompt-level regularization offers a low-cost method to counteract amplification of extreme content while retaining most personalization gains.
  • Evaluation of LLM-assisted recommenders must include measures of exposure to conspiratorial material in addition to accuracy or engagement metrics.
  • Prompt instructions themselves function as value-laden design choices that shape downstream user exposure.

Where Pith is reading between the lines

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

  • The same prompting approach could be tested on other platforms whose recommendation surfaces already contain fringe political material to check whether amplification occurs outside YouTube.
  • If statistical language regularities drive the effect, then training data composition for the LLM becomes a direct lever for controlling exposure outcomes.
  • Regulators or platforms might require disclosure of the exact prompt constraints used in production reranking so that users can anticipate ideological effects.

Load-bearing premise

That observed shifts in ideological exposure can be attributed to the LLM reranking step rather than to YouTube's existing candidate generation or to unmeasured patterns in how users select content.

What would settle it

Repeating the exact reranking procedure on the same histories but with a non-LLM baseline reranker or on a platform whose candidate pool excludes conspiratorial videos and finding no comparable increase in extreme exposure.

Figures

Figures reproduced from arXiv: 2606.02883 by Amir Ghasemian, Duncan J. Watts, Homa Hosseinmardi, Upasana Dutta.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

Recommender systems have grown from content-organization tools into sophisticated systems that shape daily behavior. By controlling what we see, they shape what we perceive, raising concerns about filter bubbles, radicalization, polarization, and social inequality. Large language models (LLMs) enable more powerful personalization, intensifying these dynamics. Yet most recommenders are tuned for engagement or limited accuracy metrics, with little attention to broader social implications, e.g. how personalization reshapes exposure in socially consequential domains. We investigate whether LLM-assisted reranking, while improving personalization, inadvertently amplifies exposure to ideologically extreme or conspiratorial political content, a risk theorized but not empirically characterized in news recommendation. Using real news-consumption histories, we rerank YouTube's sidebar candidates through zero-shot, instruction-based prompting. We compare a baseline prompt with a constrained variant that preserves topical relevance and broadens ideological exposure while reducing conspiratorial or extreme content. Without constraints, reranking strengthened personalization but increased exposure to conspiratorial and extremist material for users whose histories contained such content. Lightweight prompt-level regularization reduced promotion of extreme content and increased ideological diversity, with modest relevance loss. Synthetic experiments suggest that LLMs rerank via statistical regularities in language rather than semantic understanding of ideology, clarifying why naive prompts amplify these patterns and why regularization can reshape them. Together, our results highlight the power of LLMs to operationalize contextual nuance in high-stakes recommendation, and the need to evaluate LLM-assisted personalization beyond accuracy and treat prompt design as a value-laden rather than neutral default.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that zero-shot LLM reranking of YouTube sidebar candidates, using real user news-consumption histories, strengthens personalization but increases exposure to conspiratorial and extremist content for users whose histories already contain such material; a lightweight constrained prompt reduces extreme-content promotion, increases ideological diversity, and incurs only modest relevance loss. Synthetic experiments are presented to argue that these effects arise from statistical language regularities rather than semantic understanding of ideology.

Significance. If the attribution and measurement claims hold after addressing controls, the work would demonstrate a practical method for operationalizing nuanced, value-laden objectives (ideological diversity, reduced extremism) via prompt design in production recommender pipelines, moving beyond accuracy-only tuning. The combination of real-history experiments with synthetic controls on language statistics is a constructive approach to isolating LLM behavior.

major comments (3)
  1. [§4 and §5] The central causal claim—that observed shifts in conspiratorial/extremist exposure are produced by the LLM reranker rather than YouTube’s pre-existing candidate-generation and personalization—rests on an unisolated comparison. The setup feeds already-filtered sidebar candidates into the LLM; no non-LLM reranker baseline on the identical candidate set or de-biased pool is reported, leaving the incremental effect of the zero-shot prompt untested (see §4 real-user study and §5 synthetic experiments).
  2. [§4] Quantitative results for the real-user study (effect sizes, sample sizes, statistical tests, confidence intervals, or controls for user self-selection) are not provided in the reported findings, preventing evaluation of whether the reported increases in extreme exposure and the mitigation by regularization are reliable or practically meaningful.
  3. [§5] The synthetic experiments test language-statistical regularities but do not include an ablation that applies a non-LLM reranker (e.g., TF-IDF or embedding similarity) to the same candidate pool, so they cannot rule out that any observed ideological shift is an artifact of the candidate pool itself rather than LLM-specific prompting behavior.
minor comments (2)
  1. [§3] Clarify the exact definition and operationalization of “ideological diversity” and “conspiratorial content” (e.g., annotation protocol, inter-rater reliability) so that the diversity and extremism metrics can be reproduced.
  2. [Abstract] The abstract states directional findings without any numerical values; adding at least summary statistics (N, Δ, p-values) would improve readability even if full tables appear later.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§4 and §5] The central causal claim—that observed shifts in conspiratorial/extremist exposure are produced by the LLM reranker rather than YouTube’s pre-existing candidate-generation and personalization—rests on an unisolated comparison. The setup feeds already-filtered sidebar candidates into the LLM; no non-LLM reranker baseline on the identical candidate set or de-biased pool is reported, leaving the incremental effect of the zero-shot prompt untested (see §4 real-user study and §5 synthetic experiments).

    Authors: Our primary design compares the original YouTube ranking (non-LLM baseline) to LLM reranking on identical candidate sets, with the key contrast between unconstrained and constrained prompts. This isolates the incremental effect of prompt design. We agree an explicit non-LLM reranker ablation (e.g., embedding similarity) on the same pools would further strengthen attribution and will add this to both §4 and §5. revision: yes

  2. Referee: [§4] Quantitative results for the real-user study (effect sizes, sample sizes, statistical tests, confidence intervals, or controls for user self-selection) are not provided in the reported findings, preventing evaluation of whether the reported increases in extreme exposure and the mitigation by regularization are reliable or practically meaningful.

    Authors: We will revise §4 to explicitly report sample sizes, effect sizes, statistical tests, confidence intervals, and controls for user self-selection (e.g., via history matching). These details from our analysis will be added and highlighted. revision: yes

  3. Referee: [§5] The synthetic experiments test language-statistical regularities but do not include an ablation that applies a non-LLM reranker (e.g., TF-IDF or embedding similarity) to the same candidate pool, so they cannot rule out that any observed ideological shift is an artifact of the candidate pool itself rather than LLM-specific prompting behavior.

    Authors: The synthetic setup holds candidate pools fixed while varying prompts to isolate statistical regularities. We will add a non-LLM reranker ablation (TF-IDF and embedding similarity) on the same pools in revised §5 to confirm shifts are due to LLM prompting. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparison on held-out histories

full rationale

The paper reports an empirical study that feeds real YouTube sidebar candidates into zero-shot LLM prompts, measures exposure metrics on held-out user histories, and contrasts a baseline prompt against a regularized variant. No equations, fitted parameters, or self-citation chains are used to derive the reported effects; the outcomes are direct experimental measurements rather than quantities defined by construction from the same inputs. Synthetic experiments test language-statistical patterns but do not reduce the main claims to tautologies.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the central claim rests on unstated assumptions about the stability of LLM behavior across prompts and the validity of the chosen ideological and conspiratorial content labels.

pith-pipeline@v0.9.1-grok · 5830 in / 1177 out tokens · 21977 ms · 2026-06-28T12:23:15.990324+00:00 · methodology

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

63 extracted references · 12 canonical work pages

  1. [1]

    Deldjoo, F

    Y . Deldjoo, F. Nazary, A. Ramisa, J. Mcauley, G. Pellegrini, A. Bellogin, and T. D. Noia, A review of modern fashion rec- ommender systems, ACM Comput. Surv.56, 1 (2023)

    2023

  2. [2]

    J. Liu, P. Dolan, and E. R. Pedersen, Personalized news rec- ommendation based on click behavior, inProc. IUI(2010) pp. 31–40

    2010

  3. [3]

    Bobadilla, F

    J. Bobadilla, F. Ortega, A. Hernando, and A. Guti´errez, Recom- mender systems survey, Knowl.-Based Syst.46, 109 (2013)

    2013

  4. [4]

    Carraro and D

    D. Carraro and D. Bridge, Enhancing recommendation diver- sity by re-ranking with large language models, ACM Trans. Recomm. Syst.4, 1 (2025)

    2025

  5. [5]

    C. Gao, S. Wang, S. Li, J. Chen, X. He, W. Lei, B. Li, Y . Zhang, and P. Jiang, CIRS: Bursting filter bubbles by counterfactual 10 interactive recommender system, ACM Trans. Inf. Syst.42, 1 (2023)

    2023

  6. [6]

    O’Callaghan, D

    D. O’Callaghan, D. Greene, M. Conway, J. Carthy, and P. Cun- ningham, Down the (white) rabbit hole: The extreme right and online recommender systems, Soc. Sci. Comput. Rev.33, 459 (2015)

    2015

  7. [7]

    T. T. Nguyen, P.-M. Hui, F. M. Harper, L. Terveen, and J. A. Konstan, Exploring the filter bubble: the effect of using recom- mender systems on content diversity, inProc. WWW(2014) pp. 677–686

    2014

  8. [8]

    Cinus, M

    F. Cinus, M. Minici, C. Monti, and F. Bonchi, The effect of people recommenders on echo chambers and polarization, in Proc. ICWSM, V ol. 16 (2022) pp. 90–101

    2022

  9. [9]

    Krauth, Y

    K. Krauth, Y . Wang, and M. Jordan, Breaking feedback loops in recommender systems with causal inference, ACM Trans. Recomm. Syst.4, 1 (2025)

    2025

  10. [10]

    Aridor, R

    G. Aridor, R. Jim ´enez-Dur´an, R. Levy, and L. Song, The eco- nomics of social media, J. Econ. Lit.62, 1422–74 (2024)

    2024

  11. [11]

    Kleinberg, S

    J. Kleinberg, S. Mullainathan, and M. Raghavan, The challenge of understanding what users want: Inconsistent preferences and engagement optimization, Manag. Sci.70, 6336 (2024)

    2024

  12. [12]

    M. D. Ekstrand, A. Razi, A. Sarcevic, M. S. Pera, R. Burke, and K. L. Wright, Recommending with, not for: Co-designing recommender systems for social good, ACM Trans. Recomm. Syst. (2025)

    2025

  13. [13]

    C. Jia, M. S. Lam, M. C. Mai, J. T. Hancock, and M. S. Bern- stein, Embedding democratic values into social media ais via societal objective functions, Proc. ACM Hum.-Comput. Inter- act.8, 1 (2024)

    2024

  14. [14]

    Bernstein, A

    M. Bernstein, A. Christin, J. Hancock, T. Hashimoto, C. Jia, M. Lam, N. Meister, N. Persily, T. Piccardi, M. Saveski,et al., Embedding societal values into social media algorithms, J. On- line Trust Saf.2(2023)

    2023

  15. [15]

    C. Pei, Y . Zhang, Y . Zhang, F. Sun, X. Lin, H. Sun, J. Wu, P. Jiang, J. Ge, W. Ou,et al., Personalized re-ranking for rec- ommendation, inProc. RecSys(2019) pp. 3–11

    2019

  16. [16]

    Sonboli, F

    N. Sonboli, F. Eskandanian, R. Burke, W. Liu, and B. Mobasher, Opportunistic multi-aspect fairness through per- sonalized re-ranking, inProc. UMAP(2020) pp. 239–247

    2020

  17. [17]

    J. Gao, B. Chen, X. Zhao, W. Liu, X. Li, Y . Wang, W. Wang, H. Guo, and R. Tang, LLM4Rerank: LLM-based auto- reranking framework for recommendations, inProc. WWW (2025) pp. 228–239

    2025

  18. [18]

    Narayanan, What if algorithmic fairness is a category error?, Contemporary Debates in the Ethics of Artificial Intelligence , 77 (2026)

    A. Narayanan, What if algorithmic fairness is a category error?, Contemporary Debates in the Ethics of Artificial Intelligence , 77 (2026)

    2026

  19. [19]

    Z. Hong, Y . Wu, Z. Zhao, S. Feng, J. Ma, J. Liu, and T. Wei, Multi-objective recommendation in the era of gener- ative AI: A survey of recent progress and future prospects, arXiv:2506.16893 (2025)

    work page arXiv 2025

  20. [20]

    Zhang, R

    J. Zhang, R. Xie, Y . Hou, W. X. Zhao, L. Lin, and J.-R. Wen, Recommendation as instruction following: A large language model empowered recommendation approach, ACM Trans. Inf. Syst.43, 1 (2025)

    2025

  21. [21]

    S. Geng, S. Liu, Z. Fu, Y . Ge, and Y . Zhang, Recommendation as language processing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5), inProc. RecSys(2022) pp. 299–315

    2022

  22. [22]

    F. Yang, Z. Chen, Z. Jiang, E. Cho, X. Huang, and Y . Lu, PALR: Personalization aware LLMs for recommendation, arXiv:2305.07622 (2023)

    work page arXiv 2023

  23. [23]

    Ziegler, S

    C.-N. Ziegler, S. M. McNee, J. A. Konstan, and G. Lausen, Im- proving recommendation lists through topic diversification, in Proc. WWW(ACM, 2005) pp. 22–32

    2005

  24. [24]

    G. K. Patro, A. Biswas, N. Ganguly, K. P. Gummadi, and A. Chakraborty, FairRec: Two-sided fairness for personal- ized recommendations in two-sided platforms, inProc. WWW (ACM, 2020) pp. 1194–1204

    2020

  25. [25]

    V . Do, S. Corbett-Davies, J. Atif, and N. Usunier, Two-sided fairness in rankings via lorenz dominance, Adv. Neural Inf. Pro- cess. Syst.34, 8596 (2021)

    2021

  26. [26]

    C. Xu, S. Chen, J. Xu, W. Shen, X. Zhang, G. Wang, and Z. Dong, P-MMF: Provider max-min fairness re-ranking in rec- ommender systems, inProc. WWW(2023) pp. 3701–3711

    2023

  27. [27]

    Hosseinmardi, A

    H. Hosseinmardi, A. Ghasemian, A. Clauset, M. Mobius, D. M. Rothschild, and D. J. Watts, Examining the consumption of rad- ical content on YouTube, PNAS118, e2101967118 (2021)

    2021

  28. [28]

    Haroon, M

    M. Haroon, M. Wojcieszak, A. Chhabra, X. Liu, P. Mohapatra, and Z. Shafiq, Auditing YouTube’s recommendation system for ideologically congenial, extreme, and problematic recommen- dations, PNAS120, e2213020120 (2023)

    2023

  29. [29]

    X. Yu, M. Haroon, E. Menchen-Trevino, and M. Wojcieszak, Nudging recommendation algorithms increases news consump- tion and diversity on YouTube, PNAS nexus3, pgae518 (2024)

    2024

  30. [30]

    Jahanbakhsh, D

    F. Jahanbakhsh, D. Zhao, T. Piccardi, Z. Robertson, Z. Epstein, S. Koyejo, and M. S. Bernstein, Value alignment of social media ranking algorithms, inProc. CHI(2026) pp. 1–26

    2026

  31. [31]

    Stray, Designing recommender systems to depolarize, arXiv:2107.04953 (2021)

    J. Stray, Designing recommender systems to depolarize, arXiv:2107.04953 (2021)

    work page arXiv 2021

  32. [32]

    Y . Hou, J. Zhang, Z. Lin, H. Lu, R. Xie, J. McAuley, and W. X. Zhao, Large language models are zero-shot rankers for recom- mender systems, inProc. ECIR(Springer, 2024) pp. 364–381

    2024

  33. [33]

    El Malki, M

    O. El Malki, M. Aubin Le Qu ´er´e, A. Monroy-Hern ´andez, and M. Horta Ribeiro, Bonsai: Intentional and personalized social media feeds, inProc. CHI(2026) pp. 1–24

    2026

  34. [34]

    Carroll, A

    M. Carroll, A. Foote, K. Feng, M. Williams, A. Dragan, W. B. Knox, and S. Milli, CTRL-Rec: Controlling recommender sys- tems with natural language, arXiv:2510.12742 (2025)

    work page arXiv 2025

  35. [35]

    Milli, M

    S. Milli, M. Carroll, Y . Wang, S. Pandey, S. Zhao, and A. D. Dragan, Engagement, user satisfaction, and the amplification of divisive content on social media, PNAS nexus4, pgaf062 (2025)

    2025

  36. [36]

    E. Loru, J. Nudo, N. Di Marco, A. Santirocchi, R. Atzeni, M. Cinelli, V . Cestari, C. Rossi-Arnaud, and W. Quattrociocchi, The simulation of judgment in LLMs, PNAS122, e2518443122 (2025)

    2025

  37. [37]

    J. Kim, J. Evans, and A. Schein, Linear representations of political perspective emerge in large language models, arXiv:2503.02080 (2025)

    work page arXiv 2025

  38. [38]

    Piccardi, M

    T. Piccardi, M. Saveski, C. Jia, J. Hancock, J. L. Tsai, and M. S. Bernstein, Reranking partisan animosity in algorithmic social media feeds alters affective polarization, Science390, 10.1126/science.adu5584 (2025)

    work page doi:10.1126/science.adu5584 2025

  39. [39]

    D. Liu, B. Yang, H. Du, D. Greene, N. Hurley, A. Lawlor, R. Dong, and I. Li, RecPrompt: A self-tuning prompting frame- work for news recommendation using large language models, inProc. CIKM(2024) pp. 3902–3906

    2024

  40. [40]

    T. Wu, S. Zhang, K. Song, S. Xu, S. Zhao, R. Agrawal, S. R. In- durthi, C. Xiang, P. Mittal, and W. Zhou, Instructional segment embedding: Improving LLM safety with instruction hierarchy, arXiv:2410.09102 (2024)

    work page arXiv 2024

  41. [41]

    Raheja, N

    T. Raheja, N. Pochhi, and F. Curie, Recent advancements in LLM red-teaming: Techniques, defenses, and ethical consider- ations, arXiv:2410.09097 (2024)

    work page arXiv 2024

  42. [42]

    A. Y . Chen, B. Nyhan, J. Reifler, R. E. Robertson, and C. Wil- son, Subscriptions and external links help drive resentful users to alternative and extremist YouTube channels, Sci. Adv.9, eadd8080 (2023). 11

    2023

  43. [43]

    Hosseinmardi, A

    H. Hosseinmardi, A. Ghasemian, M. Rivera-Lanas, M. Horta Ribeiro, R. West, and D. J. Watts, Causally es- timating the effect of youtube’s recommender system using counterfactual bots, PNAS121, e2313377121 (2024)

    2024

  44. [44]

    Google, Youtube data api (v3) – videos (2026), https://developers.google.com/youtube/v3/docs/videos (Ac- cessed: 2025-03-05)

    2026

  45. [45]

    Dutta, H

    U. Dutta, H. Hosseinmardi, A. Ghasemian, A. Clauset, and D. J. Watts, Asymmetric rhetorical strategies drive partisanship in television news (2026), under review

    2026

  46. [46]

    Ghasemian, H

    A. Ghasemian, H. Hosseinmardi, U. Dutta, J. Allen, D. M. Rothschild, and D. J. Watts, News-like information ecosystem on youtube (2026), under preparation

    2026

  47. [47]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, Adv. Neural Inf. Process. Syst.33, 1877 (2020)

    2020

  48. [48]

    K. G. Vaidyanathan, M. Varun, and B. Das, Llm re-ranker as a tool for enhancing sequential recommendations, inIntell. Comput.-Proc. Comput. Conf.(Springer, 2025) pp. 20–31

    2025

  49. [49]

    Y . Luo, Q. Qin, H. Zhang, M. Cheng, R. Yan, K. Wang, and J. Ouyang, Molar: Multimodal LLMs with collaborative filtering alignment for enhanced sequential recommendation, arXiv:2412.18176 (2024)

    work page arXiv 2024

  50. [50]

    Zhang, Y

    C. Zhang, Y . Sun, J. Chen, J. Lei, M. Abdul-Mageed, S. Wang, R. Jin, S. Park, N. Yao, and B. Long, SPAR: Personalized content-based recommendation via long engagement attention, arXiv:2402.10555 (2024)

    work page arXiv 2024

  51. [51]

    Y . He, X. Liu, A. Zhang, Y . Ma, and T.-S. Chua, LLM2Rec: Large language models are powerful embedding models for se- quential recommendation, inProc. KDD(2025) pp. 896–907

    2025

  52. [52]

    Jiang, The impact of political ideology, knowledge and par- ticipation on electoral conspiracy endorsement, Front

    Y . Jiang, The impact of political ideology, knowledge and par- ticipation on electoral conspiracy endorsement, Front. Polit. Sci.5, 1069468 (2023)

    2023

  53. [53]

    A. Rao, F. Morstatter, and K. Lerman, Partisan asymmetries in exposure to misinformation, Sci. Rep.12, 15671 (2022)

    2022

  54. [54]

    F. Wu, Y . Qiao, J.-H. Chen, C. Wu, T. Qi, J. Lian, D. Liu, X. Xie, J. Gao, W. Wu,et al., MIND: A large-scale dataset for news recommendation, inProc. ACL(2020) pp. 3597–3606

    2020

  55. [55]

    A. R. KhudaBukhsh, R. Sarkar, M. S. Kamlet, and T. Mitchell, We don’t speak the same language: Interpreting polarization through machine translation, inProc. AAAI, V ol. 35 (2021) pp. 14893–14901

    2021

  56. [56]

    Immorlica, M

    N. Immorlica, M. Jagadeesan, and B. Lucier, Clickbait vs. qual- ity: How engagement-based optimization shapes the content landscape in online platforms, inProc. WWW(2024) pp. 36– 45

    2024

  57. [57]

    arXiv preprint arXiv:2501.15048 , year=

    H. Habib and R. Nithyanand, YouTube recommendations re- inforce negative emotions: Auditing algorithmic bias with emotionally-agentic sock puppets, arXiv:2501.15048 (2025)

    work page arXiv 2025

  58. [58]

    Castelo, K

    N. Castelo, K. Kushlev, A. F. Ward, M. Esterman, and P. B. Reiner, Blocking mobile internet on smartphones improves sustained attention, mental health, and subjective well-being, PNAS nexus4, pgaf017 (2025)

    2025

  59. [59]

    Quattrociocchi, V

    W. Quattrociocchi, V . Capraro, and M. Perc, Epistemolog- ical fault lines between human and artificial intelligence, arXiv:2512.19466 (2025)

    work page arXiv 2025

  60. [60]

    Adomavicius, J

    G. Adomavicius, J. Bockstedt, S. P. Curley, J. Zhang, and S. Ransbotham, The hidden side effects of recommendation systems, MIT Sloan Manag. Rev.60, 1 (2019)

    2019

  61. [61]

    Politics

    J. Rothwell, Teens spend an average of 4.8 hours on social me- dia per day (2023), accessed: 2026-05-10. 12 Appendix A: Data The data for this study comes from prior work that used Nielsen’s nationally representative desktop web panel, which records individual-level URL visits from October 2021 to De- cember 2022 [43]. Nielsen provides information only ab...

    2023

  62. [62]

    Economics, 3

    Elections and Candidates, 2. Economics, 3. Healthcare, 4. Climate and Environment, 5. Inequality and Discrimination, 6. Public Health and Safety, 7. Technology, Governance, and Public Policy, 8. Immigration, 9. Reproductive Rights, 10. Foreign Policy, 11. Firearms Policy,

  63. [63]

    topic_1": {

    Education, 13. Crime and Criminal Justice, 14. War and Conflict, 15. Political Conduct and Interpersonal Conflicts, 16. Religion. ########## OUTPUT FORMAT ########## Return in JSON format exactly: { "topic_1": { "name": "One topic from list or NA", "subtopics": ["Subtopic1", "Subtopic2"] Or ["NA"], "reason": "[Brief explanation]", "relevance score": [A nu...