Pith. sign in

REVIEW 4 major objections 4 minor 62 references

A compact on-device assistant can learn preference-aligned proactive behavior from purely synthetic interactions with 1,000 personas, then personalize per user through activation steering to match human-feedback tuning without real user log

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 04:47 UTC pith:MLFNORBH

load-bearing objection A genuinely new simulation-to-personalization pipeline, but the headline numbers measure GPT-4.1 agreeing with itself; the human study only partially rescues it. the 4 major comments →

arxiv 2602.04000 v2 pith:MLFNORBH submitted 2026-02-03 cs.HC

After Talking with 1,000 Personas: Learning Preference-Aligned Proactive Assistants From Large-Scale Persona Interactions

Ziyi Xuan , Yiwen Wu , Zhaoyang Yan , Vinod Namboodiri , Yu Yang This is my paper
classification cs.HC
keywords proactive assistantspreference alignmentpersona simulationgenerative agentson-device personalizationactivation steeringsupervised fine-tuninghuman-computer interaction
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 sets out to solve the cold-start problem for proactive smart assistants: how to be useful from the first interaction when mistimed or intrusive behavior erodes trust. Its proposal is to learn shared preference structure from a large synthetic population before any real user appears—1,000 census-aligned personas acting out week-long routines, with interactions annotated along five recurring preference dimensions (scheduling, domain priorities, autonomy, communication style, context adaptation). Stage 1 fine-tunes a compact 3B model on these synthetic traces; Stage 2 keeps that model frozen and adapts per user by steering internal activations along per-category directions derived from lightweight feedback such as accept/reject signals and satisfaction ratings. In simulation this lifts timing appropriateness (TAI) from 0.140 to 0.817 and interaction quality (IQA) from 0.797 to 0.915, and the steered model approaches Direct Preference Optimization (DPO) numbers (0.875 vs 0.896 TAI; 0.936 vs 0.938 IQA). A 30-participant study found users preferred the personalized responses in 62% of comparisons and reported higher satisfaction, trust, and comfort after a few adapted sessions; the paper argues this shows the structure learned from artificial personas transfers to real users.

Core claim

The paper's central claim is that user preferences for proactive assistance are not arbitrary: across diverse users they cluster into five recurring dimensions—scheduling preference, domain prioritization, autonomy level, communication style, and context adaptation—and populations differ mainly in how they weight and express these dimensions. If that is true, a compact model can learn the shared structure from simulated interactions alone, and personalization reduces to adjusting how strongly the model attends to each dimension. The paper implements this as a two-stage pipeline: Stage 1 supervises a frozen-backbone 3B model (via Low-Rank Adaptation) to predict active preference categories an

What carries the argument

The load-bearing object is the paper's GIDEA simulation platform, an LLM-based multi-agent environment that generates persona-grounded interaction traces: 1,000 census-aligned personas, week-long activity schedules, episodic memory, and natural-language preference annotations organized by the paper's five-category schema. On top of it sit two mechanisms. Stage 1's category-structured supervised fine-tuning (Low-Rank Adaptation on a 3B backbone) makes the model predict the active preference categories while generating responses, so internal representations become organized along the five dimensions rather than only producing fluent text. Stage 2's category-specific activation-vector steering

Load-bearing premise

The load-bearing premise is that the language model that role-plays the 1,000 personas produces interaction traces whose preference expressions and reactions faithfully mirror real users; if that fidelity breaks down, the population-level gains (TAI 0.140 to 0.817) may not transfer to real devices.

What would settle it

Run the identical two-stage pipeline in a field deployment with real users over several weeks, logging every proactive opportunity and whether it was welcomed, and compare against an untuned baseline and a direct-response baseline: if real-user TAI improvement falls materially below the simulated 0.140-to-0.817 gain, or if steering fails to beat the static population-average condition in IQA, the central claim fails; a sharper probe is the paper's own evaluator swap, where replacing one LLM judge with another drops IQA from 0.923 to 0.619—real users' judgments tracking the lower number would i

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

If this is right

  • If the simulation-to-real transfer holds, proactive assistants can be developed and tested before any real user data exists, shortening the costly and trust-risky early-deployment phase during which mistimed proactivity drives users to disable the feature.
  • On-device activation steering offers a practical alternative to RLHF for personalization: it matches DPO-level timing and interaction quality in simulation while requiring only lightweight signals (accept/reject, category ratings, optional text) and storing only small per-user steering vectors.
  • Because the population model generalizes to held-out activity contexts, the five-category preference schema—rather than the specific activity distribution—appears to be what carries the transferable structure.
  • Personalization becomes interpretable and reversible: per-category steering directions let users or the system adjust timing, autonomy, or communication style independently, and the bounded decaying strength allows a return to population defaults when feedback runs out.

Where Pith is reading between the lines

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

  • The paper's own cross-simulator table hints at the ceiling of the synthetic-transfer assumption: with a different LLM as evaluator, IQA falls from 0.923 to 0.619, so the population model may be tuned to one model's notion of appropriateness. A testable extension is to train the population prior on diverse simulators in parallel and use cross-simulator disagreement as a confidence signal for when t
  • If the five-category decomposition is roughly complete, each user's steering vectors become a compact, portable preference fingerprint—transferable across devices or apps, and potentially shareable via secure aggregation across users with similar vectors, options the paper does not explore.
  • The 62% preference rate from short text-only storyboards plausibly understates the effect on a real device, where users see notifications amid ongoing tasks; conversely, storyboards omit the annoyance of actual interruptions. A field study logging all proactive opportunities and user responses would disambiguate these two biases.
  • The paper itself cautions that CAS/PSC/TAI are diagnostic measures tied to its five-category decomposition rather than independent proxies for user benefit; whether those metrics track real-world outcomes such as feature retention or task completion is a testable question the paper leaves open.

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 / 4 minor

Summary. The paper proposes a population-to-individual learning framework for proactive assistants. A synthetic dataset is generated by simulating 1,000 census-aligned personas on the GIDEA platform over week-long activity schedules, with interactions annotated along five preference categories (scheduling, domain prioritization, autonomy, communication style, context adaptation). Stage 1 performs category-structured supervised fine-tuning of a 3B Llama model on this dataset; Stage 2 adapts to individual users through activation-vector steering driven by light interaction feedback. The authors report large improvements over untuned and direct-response baselines on simulated metrics (TAI 0.140→0.817, IQA 0.797→0.915), claim activation steering is comparable to RLHF/DPO, and present a 30-participant storyboard study in which personalized responses were preferred 62% of the time and adaptive-condition participants reported higher satisfaction, trust, and comfort.

Significance. If the claims hold, this is a practically valuable contribution: it offers a scalable, privacy-preserving way to cold-start on-device assistants without real user logs, and a lightweight per-user adaptation mechanism that avoids retraining. The paper is also honest in several places: §8.1 states that CAS/PSC/TAI are diagnostic rather than user-benefit metrics, and §8.5 warns that simulator-based results are controlled comparisons, not absolute estimates of real-user quality. Additional strengths include the scale of the synthetic dataset, the validation of activity distributions, cross-model and cross-simulator checks, and the IRB-approved human study. However, the load-bearing quantitative evidence remains inside a largely closed loop: GPT-4.1 generates the training signal and also acts as the judge for the headline metrics. The paper's own cross-simulator table (Table 5) shows that the primary outcome, IQA, drops below the untuned baseline under one alternative judge. The human study does not currently close that gap because it does not measure TAI, does not compare against DPO, and reports no significance tests for the adaptation study. These issues are fixable but require addit

major comments (4)
  1. [§8.5, Table 5 and §8.1] The primary outcome is not robust to changing the simulated user. Table 5 reports that with Claude-Sonnet-4 as the user, IQA falls to 0.619, below the untuned baseline's 0.797 from Table 3; with Llama-3.1-70B, TAI drops from 0.875 to 0.404. Since §8.1 designates IQA as the primary user-facing metric and §8.3 uses these simulated gains as evidence of improved interaction quality, the headline claims are not established by the simulation. The paper should report human IQA/TAI ratings for the same interactions, or at minimum re-run the headline comparison with evaluator-specific baselines and error bars.
  2. [§5.2.2 and §8.2] The evaluation loop is partly closed: GPT-4.1 generates the preference expressions and preferred responses used in Stage-1 training and also acts as the simulated user who provides TAI, IQA, and category labels in evaluation. Consequently, TAI/CAS/PSC largely measure agreement with GPT-4.1's own preference consensus. The authors disclose in §8.1 and §8.5 that structured metrics are diagnostic and simulation results are controlled comparisons, but the abstract and §8.3 go further and claim 'comparable to reinforcement learning from human feedback.' That specific claim currently depends on GPT-4.1 agreeing with itself. A concrete test would be to collect TAI and IQA from human raters for the steering-vs-DPO comparison, or to use a non-GPT-4.1 judge for that comparison.
  3. [§9 and §10] The human study is too weak to support the §10 claim that the learned structure 'transfers to real users' or that activation steering is sufficient for real-world personalization. The preference detection study compares personalized responses against a non-personalized baseline, not against DPO or steering; the 62% preference rate is reported without a confidence interval or significance test. The interactive adaptation study has n=10 per condition, uses a between-subjects design, reports no inferential statistics, and TAI is never measured with human raters. Please report effect sizes and significance tests, and ideally include a human-rated comparison between steering and DPO.
  4. [§6.2, §7, and Appendix A] There is a training/deployment input mismatch. Stage-1 training examples include a census-derived persona profile as an input feature, and Appendix A shows a detailed profile for P042. In real deployment, as described in §7 and the user study, no such persona profile exists for an actual user. The paper does not explain how the model is conditioned at test time when the persona field is absent. This distribution shift could materially affect the claimed cold-start benefit and should be clarified or resolved.
minor comments (4)
  1. [Abstract and §1] The abstract states the human study had 34 participants, but the full text consistently reports 30 participants (§9). Please reconcile.
  2. [§8.5, Table 4] For Qwen3-1.7B, CAS/PSC/TAI are near zero (0.024/0.021/0.113) while IQA is 0.838. This large divergence between the structured metrics and IQA is worth explicit discussion, since it suggests IQA can remain high even when preference understanding fails.
  3. [§5.3, Figure 2] The activity distribution comparison shows large deviations for health (17.2% vs 10%), cooking (14.1% vs 3%), and entertainment (10.1% vs 32%). The text explains these as assistant-relevant emphases, but a formal alignment measure (e.g., correlation or divergence) would strengthen the validation claim.
  4. [General] No artifact availability statement or link to the synthetic dataset, code, or trained adapters is provided. Releasing these would substantially help reproducibility and allow the community to test the cross-simulator concerns directly.

Circularity Check

1 steps flagged

Simulated gains largely measure GPT-4.1 self-agreement: the same model that creates the Stage-1 training targets also acts as the simulated user who supplies TAI/IQA/CAS/PSC; the human study only partially breaks this loop.

specific steps
  1. fitted input called prediction [§5.2.2 (data generation), §8.1 (metrics), §8.3 (headline results)]
    "Within GIDEA’s interaction engine, GPT-4.1 role-plays each persona at every potential proactive intervention point. Given the current activity context, the persona is prompted to reflect on whether the assistant’s action is appropriate and, if so, how it should be delivered. ... in simulation, scores reflect the simulated user’s ratings."

    The same GPT-4.1 model generates the Stage-1 training targets (preferred responses, category labels, timing judgments) and then, as the simulated user, supplies the TAI/IQA/CAS/PSC evaluation labels. The headline improvements (TAI 0.140→0.817, IQA 0.797→0.915) are therefore largely agreement of the trained assistant with the GPT-4.1 consensus that produced its training signal, not an independent measure of real-user value. The paper's own Table 5 confirms the fragility: with Llama-3.1-70B as user, TAI collapses to 0.404, and with Claude-Sonnet-4, IQA drops to 0.619, below the untuned baseline's 0.797. This is a closed loop rather than a prediction from first principles.

full rationale

The central simulated claim—that a 3B model cold-started on 1,000 personas reaches RLHF-level timing and quality—rests on a closed evaluation loop: GPT-4.1 writes the preferred responses and preference labels used for Stage-1 training, and GPT-4.1 then rates the trained assistant's TAI, IQA, CAS, and PSC. That is a quotable reduction of the 'prediction' to the evaluator that supplied the training signal, and Table 5 shows how much the numbers move when the evaluator is changed (e.g., IQA 0.923→0.619 with Claude-Sonnet-4, TAI 0.875→0.404 with Llama-3.1-70B). I do not count the GIDEA self-citations [47,48] as circular per se, because GIDEA's prior validation is peer-reviewed and externally falsifiable; however, that prior work does not validate the specific use of GPT-4.1 preference expressions as training targets. The paper is transparent about the diagnostic nature of structured metrics (§8.1) and does attempt to break the loop with cross-simulator checks and a 30-person user study. The human study is real external evidence, but it does not measure TAI, does not compare against DPO/RLHF, and its 62% preference rate is against a non-personalized baseline. Hence the headline performance is partially circular: the in-simulation improvements are genuine fits to GPT-4.1, but they are presented as evidence of generalizable, RLHF-comparable quality without a fully independent evaluation of that specific claim.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The framework's free parameters are few but underreported (tau, steering bounds/decay, baseline weights), and the core axioms are the fidelity of GPT-4.1-simulated personas and the sufficiency of the five-category schema — both only partially validated. No new physical or domain entities are introduced beyond a data-derived steering construct.

free parameters (5)
  • Dissatisfaction threshold tau = 3
    Steering triggers only when a category satisfaction rating s_j,k < tau; set to 3 'in practice' (§6.3) with no sensitivity analysis.
  • Five-category preference taxonomy C = scheduling, domain prioritization, autonomy, communication style, context adaptation
    Hand-selected taxonomy from prior HCI work and the authors' own simulated traces (§3.2, Table 1); both the training objectives and the metrics CAS/PSC/TAI are organized around it.
  • Steering strength alpha and decay schedule = bounded, decays over time (values not reported)
    Inference-time steering magnitude is bounded and decays (§6.3) but concrete values and sensitivity analyses are not given.
  • Population-average preference weights for static baseline V = from an unreported 20-participant pilot
    The static-baseline condition in the interactive user study was configured from a separate pilot whose details are not provided (§9.2).
  • Simulation scale (1,000 personas, 100 interactions, 10 periods) = 1,000 / 100 / 10
    Sizes chosen for computational feasibility (§5.2.2); the scaling analysis shows flattening beyond 100 personas (§8.4), so 1,000 is a resource choice rather than a tuned parameter.
axioms (4)
  • domain assumption GPT-4.1 role-play personas faithfully emulate real user preference expression and reactions to proactive assistance.
    Stage 1's entire training signal comes from GIDEA simulation (§5.2.2), justified by the authors' own prior GIDEA validation [47,48] and 'algorithmic fidelity' [4]. The paper's own Table 5 shows IQA falls to 0.619 with a Claude-Sonnet-4 evaluator, making this assumption fragile.
  • domain assumption The five-category preference decomposition is sufficient to capture proactive interaction quality.
    The assistant attends only to these five dimensions; metrics CAS/PSC/TAI and steering directions are defined over them (§3.3, §6.3, §8.1). Quality is therefore measured against the paper's own taxonomy, as the authors concede in §8.1.
  • domain assumption Activation steering directions generalize across contexts and do not collapse over long horizons.
    Steered activations h'_l = h_l + sum_k alpha_k v_k are applied at inference for all contexts; stability evidence comes only from simulation over 500 interactions (§8.4, Figure 7), not from users.
  • ad hoc to paper Reflexion-based episodic memory in the simulator produces realistic preference evolution rather than drift artifacts.
    Preference change over sessions is manufactured by reflexion summarization prompts (§5.2.2); no validation against how real user preferences evolve over time.
invented entities (1)
  • Per-user per-category steering direction vector v_u^(k) no independent evidence
    purpose: Latent activation-space object used to shift assistant behavior toward a user's preferences without weight updates (§6.3).
    A data-derived construct (difference of mean activations), not a falsifiable entity; its only evidence is the behavioral improvement claims inside this paper. The five-category taxonomy is treated here as an axiom rather than an entity.

pith-pipeline@v1.3.0-alltime-deepseek · 22735 in / 18639 out tokens · 169179 ms · 2026-08-03T04:47:22.216661+00:00 · methodology

0 comments
read the original abstract

Smart assistants increasingly act proactively, yet mistimed or intrusive behavior often causes users to lose trust and disable these features. Learning user preferences for proactive assistance is difficult because real-world studies are costly, limited in scale, and rarely capture how preferences change across multiple interaction sessions. Large language model based generative agents offer a way to simulate realistic interactions, but existing synthetic datasets remain limited in temporal depth, diverse personas, and multi-dimensional preferences. They also provide little support for transferring population-level insights to individual users under on-device constraints. We present a population-to-individual learning framework for preference-aligned proactive assistants that operates under on-device and privacy constraints. Our approach uses large-scale interaction simulation with 1,000 diverse personas to learn shared structure in how users express preferences across recurring dimensions such as timing, autonomy, and communication style, providing a strong cold start without relying on real user logs. The assistant then adapts to individual users on device through lightweight activation-based steering driven by simple interaction feedback, without model retraining or cloud-side updates. We evaluate the framework using controlled simulations with 1,000 simulated personas and a human-subject study with 34 participants. Results show improved timing decisions and perceived interaction quality over untuned and direct-response baselines, while on-device activation steering achieves performance comparable to reinforcement learning from human feedback. Participants also report higher satisfaction, trust, and comfort as the assistant adapts over multiple sessions of interactions.

Figures

Figures reproduced from arXiv: 2602.04000 by Vinod Namboodiri, Yiwen Wu, Yu Yang, Zhaoyang Yan, Ziyi Xuan.

Figure 1
Figure 1. Figure 1: System Overview 4 Overview [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Generated activity distribution compared to [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Customized web-based interface used in the Preference Detection Study. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Together, they show both cold-start behavior and later adaptation for seen and unseen contexts. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Boxplots of Preference Understanding (CAS/PSC), Temporal Appropriateness Index (TAI), and Interaction Quality [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Scaling of population-level learning with train [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: User perception of proactive adaptation across conditions. assistant’s decisions and actions by the end of the study?" (5-point scale from strongly decreased to strongly increased), (4) "How did the assistant’s learning or changes affect your comfort, satisfaction, or willingness to use it again?" (5-point scale from strongly decreased to strongly increased), and (5) overall satisfaction (5-point scale fro… view at source ↗
Figure 9
Figure 9. Figure 9: Example of a storyboard-based preference comparison task used in the Preference Detection Study. [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of a storyboard-based preference comparison task used in the Interactive Adaptation Study. [PITH_FULL_IMAGE:figures/full_fig_p026_10.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

62 extracted references · 3 canonical work pages

  1. [1]

    acon96. 2024. Home-Assistant-Requests. https://huggingface.co/datasets/acon96/Home-Assistant-Requests Accessed: 2025-11-16

    2024

  2. [2]

    Meta AI. 2024. Llama 3.2 3B Instruct. https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct. Accessed: 2024-11-16

    2024

  3. [3]

    Anthropic. 2024. Claude 4 Model Family. https://www.anthropic.com/claude

    2024

  4. [4]

    Lisa P Argyle, Ethan C Busby, Nancy Fulda, Joshua R Gubler, Christopher Rytting, and David Wingate. 2023. Out of one, many: Using language models to simulate human samples.Political Analysis31, 3 (2023), 337–351

    2023

  5. [5]

    Frank Bentley, Chris Luvogt, Max Silverman, Rushani Wirasinghe, Brooke White, and Danielle Lottridge. 2018. Understanding the long-term use of smart speaker assistants.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies2, 3 (2018), 1–24

    2018

  6. [6]

    Caterina Bérubé, Marcia Nißen, Rasita Vinay, Alexa Geiger, Tobias Budig, Aashish Bhandari, Catherine Rachel Pe Benito, Nathan Ibarcena, Olivia Pistolese, Pan Li, et al . 2024. Proactive behavior in voice assistants: A systematic review and conceptual model. Computers in Human Behavior Reports14 (2024), 100411

    2024

  7. [7]

    Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Iñigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gašić. 2018. Multiwoz–a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling.arXiv preprint arXiv:1810.00278(2018). 22•Ziyi Xuan, Yiwen Wu, Zhaoyang Yan, Vinod Namboodiri, and Yu Yang

    Pith/arXiv arXiv 2018

  8. [8]

    Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, et al. 2023. Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback.arXiv preprint arXiv:2307.15217(2023)

    Pith/arXiv arXiv 2023

  9. [9]

    Narae Cha, Auk Kim, Cheul Young Park, Soowon Kang, Mingyu Park, Jae-Gil Lee, Sangsu Lee, and Uichin Lee. 2020. Hello there! is now a good time to talk? Opportune moments for proactive interactions with smart speakers.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies4, 3 (2020), 1–28

    2020

  10. [10]

    Valerie Chen, Alan Zhu, Sebastian Zhao, Hussein Mozannar, David Sontag, and Ameet Talwalkar. 2024. Need Help? Designing Proactive AI Assistants for Programming. arXiv:2410.04596 [cs.HC] https://arxiv.org/abs/2410.04596

    Pith/arXiv arXiv 2024

  11. [11]

    Jessie Chin, Smit Desai, Sheny Lin, and Shannon Mejia. 2024. Like my aunt dorothy: effects of conversational styles on perceptions, acceptance and metaphorical descriptions of voice assistants during later adulthood.Proceedings of the ACM on Human-Computer Interaction8, CSCW1 (2024), 1–21

    2024

  12. [12]

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences.Advances in neural information processing systems30 (2017)

    2017

  13. [13]

    Leigh Clark, Nadia Pantidi, Orla Cooney, Philip Doyle, Diego Garaialde, Justin Edwards, Brendan Spillane, Emer Gilmartin, Christine Murad, Cosmin Munteanu, Vincent Wade, and Benjamin R. Cowan. 2019. What makes a good conversation?: Challenges in designing truly conversational agents. InProceedings of the 2019 CHI Conference on Human Factors in Computing S...

    arXiv 2019

  14. [14]

    Diane J Cook, Aaron S Crandall, Brian L Thomas, and Narayanan C Krishnan. 2012. CASAS: A smart home in a box.Computer46, 7 (2012), 62–69

    2012

  15. [15]

    Alshammari et al. 2017. Openshs: Open smart home simulator.Sensors17, 5 (2017), 1003

    2017

  16. [16]

    Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. 2020. Personalized Federated Learning: A Meta-Learning Approach. arXiv:2002.07948 [cs.LG] https://arxiv.org/abs/2002.07948

    Pith/arXiv arXiv 2020

  17. [17]

    Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. 2024. Large language models empowered agent-based modeling and simulation: A survey and perspectives.Humanities and Social Sciences Communications11, 1 (2024), 1–24

    2024

  18. [18]

    Diana Gavilan and Gema Martinez-Navarro. 2022. Exploring user’s experience of push notifications: a grounded theory approach. Qualitative Market Research: An International Journal25, 2 (2022), 233–255

    2022

  19. [19]

    Shubham Ghosh, Abhinav Mehrotra, and Mirco Musolesi. 2023. Privacy-Aware Context Sensing on Mobile Devices.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies7, 3, Article 127 (2023), 30 pages. doi:10.1145/3610907

    doi:10.1145/3610907 2023

  20. [20]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

    Pith/arXiv arXiv 2024

  21. [21]

    Eric Horvitz. 1999. Principles of mixed-initiative user interfaces. InProceedings of the SIGCHI conference on Human Factors in Computing Systems

    1999

  22. [22]

    Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu. 2023. Evaluating and Inducing Personality in Pre-trained Language Models. arXiv:2206.07550 [cs.CL] https://arxiv.org/abs/2206.07550

    Pith/arXiv arXiv 2023

  23. [23]

    Andreas Köpf, Yannic Kilcher, Dimitri Von Rütte, Sotiris Anagnostidis, Zhi Rui Tam, Keith Stevens, Abdullah Barhoum, Duc Nguyen, Oliver Stanley, Richárd Nagyfi, et al. 2023. Openassistant conversations-democratizing large language model alignment.Advances in neural information processing systems36 (2023), 47669–47681

    2023

  24. [24]

    Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. arXiv:2306.03341 [cs.LG] https://arxiv.org/abs/2306.03341

    Pith/arXiv arXiv 2024

  25. [25]

    Xiaoyi Liu, Yingtian Shi, Chun Yu, Cheng Gao, Tianao Yang, Chen Liang, and Yuanchun Shi. 2023. Understanding in-situ programming for smart home automation. 7, 2 (2023), 1–31. doi:10.1145/3596254

    work page doi:10.1145/3596254 2023

  26. [26]

    Yaxi Lu, Shenzhi Yang, Cheng Qian, Guirong Chen, Qinyu Luo, Yesai Wu, Huadong Wang, Xin Cong, Zhong Zhang, Yankai Lin, et al

  27. [27]

    Yuxuan Lu, Bingsheng Yao, Hansu Gu, Jing Huang, Zheshen Jessie Wang, Yang Li, Jiri Gesi, Qi He, Toby Jia-Jun Li, and Dakuo Wang

  28. [28]

    2024.Llama-3.1-70B-Instruct Model Documentation

    Meta. 2024.Llama-3.1-70B-Instruct Model Documentation. Hugging Face. https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct

    2024

  29. [29]

    Alexander Miller, Will Feng, Dhruv Batra, Antoine Bordes, Adam Fisch, Jiasen Lu, Devi Parikh, and Jason Weston. 2017. Parlai: A dialog research software platform. InProceedings of the 2017 conference on empirical methods in natural language processing: system demonstrations. 79–84

    2017

  30. [30]

    Clifford Nass, Jonathan Steuer, and Ellen R Tauber. 1994. Computers are social actors.Proceedings of the SIGCHI conference on Human factors in computing systems(1994)

    1994

  31. [31]

    nvidia. 2025. Nemotron-Dataset-USA. https://huggingface.co/datasets/nvidia/Nemotron-Personas Accessed: 2025-11-16

    2025

  32. [33]

    Jeesun Oh, Wooseok Kim, Sungbae Kim, Hyeonjeong Im, and Sangsu Lee. 2024. Better to Ask Than Assume: Proactive Voice Assistants’ Communication Strategies that Respect User Agency in a Smart Home Environment. InProceedings of the CHI Conference on Human Factors in Computing Systems. ACM, 1–17. doi:10.1145/3613904.3642193

    arXiv 2024

  33. [34]

    OpenAI. 2025. GPT-4.1 System Card. https://openai.com/index/gpt-4-1/ Accessed: February 5, 2026

    2025

  34. [35]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in neural information processing systems35 (2022), 27730–27744

    2022

  35. [36]

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology. 1–22

    2023

  36. [37]

    Joon Sung Park, Carolyn Q Zou, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Robb Willer, Percy Liang, and Michael S Bernstein. 2024. Generative agent simulations of 1,000 people.arXiv preprint arXiv:2411.10109(2024)

    Pith/arXiv arXiv 2024

  37. [38]

    Martin Pielot, Amalia Vradi, and Souneil Park. 2018. Dismissed! a detailed exploration of how mobile phone users handle push notifications. InProceedings of the 20th international conference on human-computer interaction with mobile devices and services. 1–11

    2018

  38. [39]

    Xavier et al. Puig. 2018. Virtualhome: Simulating household activities via programs. InProceedings of the IEEE conference on computer vision and pattern recognition. 8494–8502

    2018

  39. [40]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems36 (2023), 53728–53741

    2023

  40. [42]

    Silvia Schiaffino and Analía Amandi. 2004. User–interface agent interaction: personalization issues.International Journal of Human- Computer Studies60, 1 (2004), 129–148. doi:10.1016/j.ijhcs.2003.09.003

    work page doi:10.1016/j.ijhcs.2003.09.003 2004

  41. [43]

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv:2303.11366 [cs.AI] https://arxiv.org/abs/2303.11366

    Pith/arXiv arXiv 2023

  42. [44]

    Irina Shklovski, Janet Vertesi, and Dave Randall. 2014. Unpacking the Sensing Work of Mobile Systems. InProceedings of the CHI Conference on Human Factors in Computing Systems. ACM, 393–402. doi:10.1145/2556288.2557029

    arXiv 2014

  43. [45]

    Vazquez, Ulisse Mini, and Monte MacDiarmid

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2024. Steering Language Models With Activation Engineering. arXiv:2308.10248 [cs.CL] https://arxiv.org/abs/2308.10248

    Pith/arXiv arXiv 2024

  44. [46]

    Jing Wei, Tilman Dingler, and Vassilis Kostakos. 2021. Understanding user perceptions of proactive smart speakers.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies5, 4 (2021), 1–28

    2021

  45. [47]

    Ziyi Xuan, Yiwen Wu, Xuhai Xu, Vinod Namboodiri, Mooi Choo Chuah, and Yu Yang. 2025. Design and Evaluation of Generative Agent-based Platform for Human-Assistant Interaction Research: A Tale of 10 User Studies.Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.9, 4, Article 229 (Dec. 2025), 46 pages. doi:10.1145/3770661

    work page doi:10.1145/3770661 2025

  46. [48]

    Ziyi Xuan, Yiwen Wu, and Yu Yang. 2025. Demo Abstract: GIDEA: Generative AI-Powered Interactive Design and Evaluation Platform for Assistant Agent Research. InProceedings of the 23rd ACM Conference on Embedded Networked Sensor Systems. 704–705

    2025

  47. [49]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al

  48. [50]

    Bufang Yang, Yunqi Guo, Lilin Xu, Zhenyu Yan, Hongkai Chen, Guoliang Xing, and Xiaofan Jiang. 2025. SocialMind: LLM-based Proactive AR Social Assistive System with Human-like Perception for In-situ Live Interactions.Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.9, 1, Article 23 (March 2025), 30 pages. doi:10.1145/3712286

    doi:10.1145/3712286 2025

  49. [51]

    Bufang Yang, Yunqi Guo, Lilin Xu, Zhenyu Yan, Hongkai Chen, Guoliang Xing, and Xiaofan Jiang. 2025. Socialmind: Llm-based proactive ar social assistive system with human-like perception for in-situ live interactions.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies9, 1 (2025), 1–30

    2025

  50. [52]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  51. [53]

    Bufang Yang, Lilin Xu, Liekang Zeng, Kaiwei Liu, Siyang Jiang, Wenrui Lu, Hongkai Chen, Xiaofan Jiang, Guoliang Xing, and Zhenyu Yan. 2025. ContextAgent: Context-Aware Proactive LLM Agents with Open-World Sensory Perceptions.arXiv preprint arXiv:2505.14668 (2025)

    arXiv 2025

  52. [54]

    Ziqi Yang, Xuhai Xu, Bingsheng Yao, Ethan Rogers, Shao Zhang, Stephen Intille, Nawar Shara, Guodong Gordon Gao, and Dakuo Wang

  53. [55]

    Bufang Yang, Lilin Xu, Liekang Zeng, Yunqi Guo, Siyang Jiang, Wenrui Lu, Kaiwei Liu, Hancheng Xiang, Xiaofan Jiang, Guoliang Xing, et al. 2025. ProAgent: Harnessing On-Demand Sensory Contexts for Proactive LLM Agent Systems.arXiv preprint arXiv:2512.06721 (2025)

    Pith/arXiv arXiv 2025

  54. [57]

    Nima Zargham, Leon Reicherts, Michael Bonfert, Sarah Theres Voelkel, Johannes Schoening, Rainer Malaka, and Yvonne Rogers. 2022. Understanding Circumstances for Desirable Proactive Behaviour of Voice Assistants: The Proactivity Dilemma. InProceedings of the 4th Conference on Conversational User Interfaces. ACM, 1–14. doi:10.1145/3543829.3543834

    arXiv 2022

  55. [58]

    Talk2care: An llm-based voice assistant for communication between healthcare providers and older adults.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies8, 2 (2024), 1–35

    2024

  56. [59]

    I want it that way

    Nima Zargham, Dmitry Alexandrovsky, Jan Erich, Nina Wenig, and Rainer Malaka. 2022. “I want it that way”: Exploring users’ customization and personalization preferences for home assistants. InCHI conference on human factors in computing systems extended 24•Ziyi Xuan, Yiwen Wu, Zhaoyang Yan, Vinod Namboodiri, and Yu Yang abstracts. 1–8

    2022

  57. [60]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric P Xing, et al. 2023. Lmsys-chat-1m: A large-scale real-world llm conversation dataset.arXiv preprint arXiv:2309.11998(2023)

    Pith/arXiv arXiv 2023

  58. [61]

    Minjun Zhu, Yixuan Weng, Linyi Yang, and Yue Zhang. 2025. Personality Alignment of Large Language Models. arXiv:2408.11779 [cs.CL] https://arxiv.org/abs/2408.11779 Learning Preference-Aligned Proactive Assistants From Large-Scale Persona Interactions•25 A End-to-End Training Example of Synthetic Preference Supervision An Example of Synthetic Preference (1...

    Pith/arXiv arXiv 2025

  59. [62]

    Xiao Zhan, Noura Abdi, William Seymour, and Jose Such. 2024. Healthcare voice AI assistants: factors influencing trust and intention to use.Proceedings of the ACM on Human-Computer Interaction8, CSCW1 (2024), 1–37

    2024

  60. [63]

    Siyan Zhao, Mingyi Hong, Yang Liu, Devamanyu Hazarika, and Kaixiang Lin. 2025. Do llms recognize your preferences? evaluating personalized preference following in llms.arXiv preprint arXiv:2502.09597(2025)

    Pith/arXiv arXiv 2025

  61. [2024]

    Proactive agent: Shifting llm agents from reactive responses to active assistance.arXiv preprint arXiv:2410.12361(2024)

    Pith/arXiv arXiv 2024

  62. [2025]

    InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems

    Uxagent: An llm agent-based usability testing framework for web design. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. 1–12