Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Steerable Chatbots: Exploring Personalization Control Interfaces via LLM Activation Steering

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Activation steering lets a user tune an LLM's output along a preference axis, and steered chatbots matched hidden preferences where prompting did not.

desk verdict The interface work is genuinely interesting, but the headline 'steering beats prompting' relies on a correlation analysis that ignores participant clustering and never compares the two correlations directly. read the letter →

arxiv 2505.04260 v3 pith:SDTXAZBQ submitted 2025-05-07 cs.HC cs.AI

Jessica Y. Bo , Tianyu Xu , Ishan Chatterjee , Katrina Passarella-Ward , Achin Kulshrestha , D Shin This is my paper · ORCID
classification cs.HCcs.AI
keywords LLMpersonalizationactivationsteeringsteerablechatbotspreferenceexpressioncold-startchatbotinterfacesuserstudysentiment-basedlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large language models default to the average user, so personalizing them usually means prompting, which cold-start users struggle to do well. This paper proposes making one internal dial visible: activation steering, where a single linear strength factor added to the model's hidden activations amplifies or dampens a preference such as budget-versus-luxury. Across five open-source LLMs, steered outputs track the applied strength nearly linearly within a per-model functional range, and steering stacks with prompting and with multiple preference dimensions. In a within-subjects user study, steered chatbots' expressed content correlated with participants' self-reported true preferences (r=0.38, p<0.001 across all steered conditions) while a prompt-only baseline showed no correlation (r=0.03, p=0.84). The paper's case is that steering can replace part of the burden of prompt specification and gives users a concrete, controllable handle on personalization.

What carries the argument

The load-bearing mechanism is a preference steering vector: a logistic-regression probe is trained to separate positive from negative exemplars of a trait inside the model's own activations, and the probe's coefficients become the vector $v_i$ that is added to the hidden state as $h_{\text{steered}} = h + d_i v_i$, where $d_i$ is the user-facing strength and negative values steer toward the opposite trait. To measure whether a reply expresses a preference, the paper computes the relative cosine similarity $\cos(e_o, \bar e_+) - \cos(e_o, \bar e_-)$ between the reply embedding and the mean embeddings of filtered Yelp review corpora for the two traits; this number is the dependent variable in most of the computational and user-study analyses. For the LEARN interface, a second mechanism updates the scalar from user feedback via $d^*_{t+1} \leftarrow d^*_t + p(\text{dissatisfaction}(x_t) \cdot \text{direction}(x_t))$, using a pretrained sentiment classifier for the dissatisfaction term and reference-phrase similarity for the direction term.

What would settle it

Have human raters read pairs of chatbot replies that differ in steering strength and rank which reply expresses, say, budget versus luxury more strongly, then compare those rankings with the relative-cosine-similarity metric used in the paper. If the metric's ordering disagrees with raters on a substantial fraction of pairs, the paper's quantitative claims about alignment with true preferences lose their support; a second falsifier would be re-running the user study with a different embedding model or different trait corpora and checking whether the correlations survive.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that soft user preferences can be operationalized as a scalar on a precomputed direction in the LLM's activation space, and that injecting that direction into the residual stream yields outputs that express the preference in proportion to the scalar. The computational experiments show a near-linear relationship between steering strength and a perplexity-normalized preference effect within a per-model functional range; they also show that prompting shifts but does not erase the steered baseline, and that two preference dimensions can be steered additively when they are not too collinear. The user study then reports that, aggregated across the SELECT, CALIBRATE, and LEARN interfaces, steered chatbots' expressed preferences correlate with participants' self-reported ground truth at r=0.38 (p<0.001), versus r=0.03 (p=0.84) for the PROMPT baseline, and the authors therefore claim steering achieves significant improvement in generating relevant preference content over prompting alone. The study simultaneously documents that participants disagree sharply about which interface they want, revealing heterogeneous values around control, transparency, effort, and persistence.

Load-bearing premise

The load-bearing premise is that the metric used to score how strongly a preference appears in a reply—closeness of the reply's embedding to the average embedding of Yelp reviews for that trait—actually matches how strongly a person would perceive the preference, since every correlation result in the paper is computed against that metric.

Editorial extensions

If this is right

  • Cold-start personalization no longer requires users to articulate preferences: one scalar per dimension can be set directly or learned within a few turns of conversation.
  • Steering and prompting are complements rather than competitors: the vector sets a latent baseline, and natural-language follow-ups add contextual variation on top of it.
  • Multiple preference dimensions can be combined into a single additive profile, at least when the dimensions are reasonably orthogonal, opening a path from single-dial demos to multi-axis assistants.
  • Interface choice is not neutral: because users split over control, transparency, and effort, a deployable steerable chatbot would likely need more than one interaction modality, or an easy way to switch.
  • Steering's inference-time-only cost makes the approach practical for on-device assistants and avoids storing an explicit user profile, which the paper argues is a privacy advantage over history-based personalization.

Reading between the lines

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

  • If the preference-effect metric is valid, the near-linear strength-response curves suggest a calibration protocol: a few pairwise comparisons could place a user on each dimension's monotone segment and map their answer to a steering strength, which is essentially what CALIBRATE approximates with only two or three rounds.
  • The per-model functional range implies that the same numeric dial means different things on different models; an interface that remaps slider positions to each model's linear zone could outperform the fixed -100 to 100 mapping used here.
  • The user study's heterogeneity hints that a combined interface—direct slider for immediate control, a calibration pass for initialization, and transparent learned updates for persistence—might satisfy more users than any single design, though the paper only suggests this as future work.
  • A direct test of the framework would be whether steering strengths inferred from one conversation predict a user's choices in a new task domain; if they do, the scalar profile is a genuine preference model rather than a per-task fitting artifact.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes activation steering as a lightweight method for personalizing LLM responses along interpretable preference dimensions, exposing the steering strength as a linear factor that users can control directly. It constructs steering vectors for five preference dimensions using GPT-4o-generated contrastive samples, and evaluates the resulting control through four computational experiments (E1-E4) across five open-source LLMs, measuring expressed preference as relative cosine similarity to Yelp review corpus embeddings. It then reports a within-subjects user study (n=14) comparing three steering-based interfaces (SELECT, CALIBRATE, LEARN) against a prompting-only baseline, with quantitative analyses of steering-effect correspondence, alignment with self-reported ground-truth preferences, subjective perceptions, and qualitative interview themes. The central claim is that steered chatbots align with users' underlying preferences significantly better than prompting alone, while survey and interview data reveal heterogeneous preferences among users regarding control, transparency, and persistence.

Significance. If the central measurement and statistical claims held, this would be a useful contribution: it provides a resource-efficient, cold-start personalization mechanism, a broad computational exploration across five models and five preference dimensions, and a genuinely user-centered comparison of three interaction designs for exposing a linear steering factor. The computational dose-response results in E1 are internally consistent across models and dimensions, and the qualitative analysis in U4 is informative and well-grounded in participant quotes. The manuscript is also honest about several limitations of the prototype learning and calibration algorithms. However, the headline empirical claim rests on a message-level correlation analysis that ignores participant nesting, and the core metric for 'expressed preference' is never validated against human judgments. These are load-bearing issues that require reanalysis and additional validation rather than mere editing. The paper does not release code, data, or analysis scripts, which limits independent verification.

major comments (3)
  1. [Section 5.3, Figure 8b] The U2 analysis treats each LLM message as an independent observation, but the predictor (ground-truth preference) is a single self-report per participant per task, repeated for every message from that participant. With n=14 participants, the effective sample size is far smaller than the number of plotted points, so the reported p<0.001 for All Steering does not by itself support the headline claim that steering outperforms prompting. The authors should reanalyze at the participant level (or with a mixed-effects model including random intercepts for participants) and, crucially, report a direct test of whether the All Steering correlation (r=0.38) is significantly higher than the PROMPT correlation (r=0.03); separate p-values do not establish a difference between two correlations.
  2. [Section 4.2] The central measurement premise, expressed preference as Effect = cos(e_o, \bar e_+) - cos(e_o, \bar e_-) against mean Yelp review embeddings, is never validated against human judgments. Every quantitative claim in E1-E4 and U1-U2 depends on this operationalization, but the paper provides no evidence that relative cosine similarity to mean review embeddings tracks how strongly a preference is expressed in text, as opposed to topic or lexical overlap. A validation study with human ratings of preference expression on the same steered outputs, or use of an existing labeled preference-strength corpus, is needed before the dose-response and user-alignment claims can be interpreted.
  3. [Section 5.3, U3 analysis] The 'Favourite' comparison is circular: the favourite interface is defined as the one with the highest Likelihood-to-Use rating, and that same outcome is then compared between Favourite and PROMPT. This guarantees by construction that Favourite scores at least as high as PROMPT on Likelihood to Use, so the reported 'highly significant differences across all perception categories' do not constitute evidence that steerable chatbots are preferred. The authors should define favourite from an independent measure (for example, an explicit preference ranking question) or otherwise adjust the analysis for the selection rule.
minor comments (4)
  1. [Section 4.3, Figure 5, Appendix F] The computational experiment is introduced as E4 in Section 4.3 but is labelled E5 in Figure 5 and Appendix F; the numbering should be harmonized throughout.
  2. [Equation (1)] The update rule is underspecified because p(·) is only described as a 'simple linear transformation' and the functional steering range used by p is determined post hoc per model in Appendix A; the exact mapping and the criterion used to set the range should be stated explicitly.
  3. [Figure 8] The scatterplots plot many overlapping points, making the reported correlations difficult to assess visually; adding transparency or jitter and reporting participant-level summary statistics would improve transparency.
  4. [Throughout] There are minor typographical errors ('jewlery' in Table 1, 'disatisfied', 'feedback feedback', 'althougth'), and the manuscript does not state a data/code availability plan, which would help independent verification of the computational experiments.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity; the central steering-vs-prompting claim rests on independent data, with one E4 simulation step that reduces to reading explicit feedback.

  1. self definitional [Section 3 (Equation 1 and direction(·) definition) with Section 4.4 E4 and Appendix G Listings G.1-G.2]
    "GPT is prompted with in-context examples of what the user's preference at these levels, and is instructed to output a response to the LLM's output with the 1) level of satisfaction that the user would feel, and 2) the direction of change (cheaper or more luxurious) that the user would want. ... direction(·) is determined based on the greater cosine similarity to Sentence-BERT embeddings of reference phrases (e.g., “I want more luxury” and “I want lower cost” for the preference dimension of cost)."

    In E4 the simulated user's 'hidden' preference h is not hidden from the learning signal: the GPT-as-user is instructed to state the direction of change (cheaper/more luxurious), and the algorithm's direction(·) detects that direction by cosine similarity to reference phrases that the generated message is told to embed. The learned steering therefore converges to h because h is directly encoded in each feedback message, so the experiment demonstrates that the update rule can read an explicit directional hint rather than inferring a latent preference. The paper itself limits this to a precursor validation, and the central user-study comparison (U2) is independent of this step.

full rationale

The paper's main claim that steered chatbots align better with underlying user preferences than prompting alone is grounded in a non-circular chain: steering vectors are trained on GPT-4o-generated gold-standard examples, while preference expression is measured against human-written Yelp review corpora, and the user-study ground truth is a self-report collected before the tasks. The U2 correlation between expressed preference and self-reported ground truth is empirical, not forced by construction, and the PROMPT baseline's near-zero correlation shows that the steering effect is not an artifact of the metric. The per-model functional steering range is calibrated in E1 and reused downstream, but the paper defines that range via perplexity degeneration rather than via the linearity claim, so the 'near-linear within functional range' finding is not tautological. There are no load-bearing self-citations: the steering methodology is attributed to external prior work (von Rütte et al.), and the evaluation data are external. The only notable circular-adjacent step is the E4 simulated-preference learning experiment, where the simulated user is explicitly instructed to embed the target direction and the direction detector matches exactly those reference phrases; this makes the 'hidden preference learning' demonstration largely self-definitional. However, the paper explicitly labels E4 as a precursor validation, and the real-user LEARN results are reported honestly as mixed (r=0.14, p=0.20), so this step does not support the paper's central conclusion. Statistical concerns about participant clustering in U2 are validity issues rather than circularity and do not raise the circularity score.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical or model-level entities: it applies existing activation-steering vectors to preference dimensions and builds three interface designs around the scalar steering factor. The load-bearing assumptions are the representability of preferences as scalars, the validity of GPT-4o-generated steering data, and the validity of embedding-similarity-based evaluation, none of which are independently verified.

free parameters (5)
  • Functional steering range per model = stablelm (-10,10); gemma-2-2b (-30,30); Mistral-7B (-30,30); Qwen2.5-7B (-10,10); gemma-2-9b (-30,30)
    Selected by inspecting where steering degrades output (perplexity). All later experiments and the user study operate only inside this range, so the linear-effect claims are conditional on this choice.
  • p() remapping in Eq. 1 (LEARN update) = Not specified in the paper
    The update rule uses an unspecified linear transformation to convert sentiment and direction into steering units; convergence results depend on this mapping.
  • Top-k steering layers per model = 16, 16, 24, 24, 32
    Chosen per model based on probe detection accuracy; affects the strength and quality of steering.
  • Sentiment weighting for dissatisfaction = 0.75 * p_negative + 0.25 * p_neutral
    Hand-picked combination of TweetEval probabilities that defines the step size in LEARN.
  • Initial calibration extremes and rounds = d_A=-100, d_B=100, 2-3 rounds
    CALIBRATE's convergence behavior and final steering strength depend on these ad hoc choices.
assumptions (6)
  • domain assumption Each preference dimension is representable as a single continuous scalar, and the preferred response is o = M(x, steer(d_u)).
    Section 3 framework: this reduces a user's soft preferences to one number per dimension and ignores multi-factorial composition beyond additive steering.
  • domain assumption A linear direction in the residual stream, recovered by a logistic-regression probe, encodes each preference concept.
    Section 3 'Constructing Steering Vectors', following von Rütte et al.; if probes do not align with human preference semantics, the whole method fails.
  • domain assumption GPT-4o-generated examples (50-80 per trait) are valid gold-standard exemplars of the preference traits.
    Section 3 and Appendix A: the steering vectors are computed from these synthetic examples.
  • domain assumption Relative cosine similarity to mean BERT embeddings of filtered Yelp review corpora measures expressed preference strength.
    Section 4.2: all quantitative preference effects and the user study correlations use this metric; no human validation is provided.
  • domain assumption Self-reported pre-task preference on a 1-10 scale is the user's ground truth preference.
    Section 5.1 and U2: alignment is computed against these ratings, even though participants later reported changing their minds (Section 5.3, U4 'Persistency').
  • domain assumption A Twitter sentiment classifier transfers to conversational feedback in lifestyle-planning tasks.
    LEARN uses TweetEval probabilities (Section 3); no validation on this domain is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Steerable Chatbots: Exploring Personalization Control Interfaces via LLM Activation Steering." pith.science (2026). https://pith.science/paper/SDTXAZBQ

@misc{pith2026250504260,
  author       = {Pith},
  title        = {Pith review of: Steerable Chatbots: Exploring Personalization Control Interfaces via LLM Activation Steering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDTXAZBQ}},
  note         = {Machine review of arXiv:2505.04260}
}
read the original abstract

Personalizing LLM responses typically requires users to articulate their preferences through prompting, which can be burdensome at cold start and difficult to articulate in natural language. We introduce an alternative paradigm, steerable chatbots: rather than asking users to describe what they want, let them directly manipulate it via a linear factor. We implement this through activation steering, leveraging a linear scalar to control how strongly a preference is expressed in the LLM's output. We first assess the computational viability of activation steering as a method to control granular preference expression, then we explore how the factor can be exposed to users. We prototype three activation steering interface designs that vary on the axes of agency (user-led vs. system-driven) and fluidity (static vs. adaptive). A within-subjects user study (n=14) in cold-start personalization tasks shows the potential for steerable chatbots to align better with underlying user preferences than prompting alone, while revealing heterogeneous values around control, persistence, and transparency in LLM personalization.

Figures

Figures reproduced from arXiv: 2505.04260 by the authors.

Figure 1
Figure 1. We apply preference-based activation steering to improve the personalization of LLMs to better match the underlying [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Effect of expressed preferences (top) and perplexity-normalized effect (bottom) for all preferences and models for E1. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Effect of additional prompting on steering for two preferences: cost (top) and age (bottom) for E2. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Preference effects for culture and age when both [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Overview of the within-subjects user study procedure, where participants completed a different personalization task [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Screen captures of all four chatbot interfaces that participants used. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Correlation analyses with the expressed preferences in the user study conversations. These answer the questions of [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Statistics on how well each steerable chatbot [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 11
Figure 11. Figure 11: Post-experiment perceptions results for all four interfaces. Due to high heterogeneity in the participant’s preferred [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: An abbreviated user conversation for the [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Inverted Detection and Control in Steering Vectors

    cs.LG 2026-08 conditional novelty 6.0 of 10

    Highly discriminative steering vectors can reliably induce the opposite of the intended concept expression, and a representation-based spoof score can detect and correct this inversion.

Reference graph

Works this paper leans on

89 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. 2019. Gradio: Hassle-free sharing and testing of ml models in the wild. arXiv preprint arXiv:1906.02569 (2019)

    arXiv 2019

  2. [2]

    Christopher M Ackerman. 2024. Representation Tuning. arXiv [cs.LG] (Sept. 2024)

    2024

  3. [3]

    Jae-wook Ahn, Peter Brusilovsky, Jonathan Grady, Daqing He, and Sue Yeon Syn

  4. [4]

    Rumi A Allbert and James K Wiles. 2024. Identifying and manipulating personality traits in LLMs through activation engineering. arXiv [cs.CL] (Dec. 2024)

    2024

  5. [5]

    Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. arXiv [cs.LG] (June 2024)

    2024

  6. [6]

    Fedor Bakalov, Marie-Jean Meurs, Birgitta König-Ries, Bahar Sateli, René Witte, Greg Butler, and Adrian Tsang. 2013. An approach to controlling user models and personalization effects in recommender systems. In Proceedings of the 2013 international conference on Intelligent user interfaces . 49–56

    2013

  7. [7]

    F Barbieri, J Camacho-Collados, L Neves, and L Tweeteval Espinosa-Anke. 2020. Unified benchmark and comparative evaluation for tweet classification. arXiv

    2020

  8. [8]

    Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2022. Discovering latent knowledge in language models without supervision. arXiv [cs.CL] (Dec. 2022)

    2022

Show all 89 references
  1. [9]

    Yuanpu Cao, Tianrong Zhang, Bochuan Cao, Ziyi Yin, Lu Lin, Fenglong Ma, and Jinghui Chen. 2024. Personalized steering of Large Language Models: Versatile steering vectors through bi-directional preference optimization. arXiv [cs.CL] (May 2024)

    2024

  2. [10]

    Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. 2023. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint a...

    2023 arXiv

  3. [11]

    Chih-Ming Chen, Hahn-Ming Lee, and Ya-Hui Chen. 2005. Personalized e- learning system using item response theory. Computers & Education 44, 3 (2005), 237–255

    2005

  4. [12]

    Daiwei Chen, Yi Chen, Aniket Rege, Zhi Wang, and Ramya Korlakai Vinayak. [n. d.]. PAL: Sample-Efficient Personalized Reward Modeling for Pluralistic Align- ment. In The Thirteenth International Conference on Learning Representations

  5. [13]

    Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, Aili Chen, Nianqi Li, Lida Chen, Caiyu Hu, Siye Wu, Scott Ren, Ziquan Fu, and Yanghua Xiao. 2024. From persona to personalization: A survey on role-Playing Langu...

    2024

  6. [14]

    Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu

  7. [15]

    John Joon Young Chung and Eytan Adar. 2023. Artinter: AI-powered Boundary Objects for Commissioning Visual Arts. InProceedings of the 2023 ACM Designing Interactive Systems Conference. 1997–2018

    2023

  8. [16]

    John Joon Young Chung and Eytan Adar. 2023. Promptpaint: Steering text-to- image generation through paint medium-like interactions. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology . 1–17

    2023

  9. [17]

    Victoria Clarke and Virginia Braun. 2017. Thematic analysis. The journal of positive psychology 12, 3 (2017), 297–298

    2017

  10. [18]

    Hai Dang, Lukas Mecke, and Daniel Buschek. 2022. Ganslider: How users control generative models for images using multiple sliders with and without feedforward information. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems. 1–15

    2022

  11. [19]

    Bofei Gao, Feifan Song, Yibo Miao, Zefan Cai, Zhe Yang, Liang Chen, Helan Hu, Runxin Xu, Qingxiu Dong, Ce Zheng, Shanghaoran Quan, Wen Xiao, Ge Zhang, Daoguang Zan, Keming Lu, Bowen Yu, Dayiheng Liu, Zeyu Cui, Jian Yang, Lei Sha, Houfeng Wang, Zhifang Sui, Peiyi Wang, Tianyu L...

  12. [20]

    Ge Gao, Alexey Taymanov, Eduardo Salinas, Paul Mineiro, and Dipendra Misra

  13. [21]

    Andrew D Gershoff, Ashesh Mukherjee, and Anirban Mukhopadhyay. 2008. What’s not to like? Preference asymmetry in the false consensus effect. Journal of Consumer Research 35, 1 (2008), 119–125

    2008

  14. [22]

    Davor Hafnar and Jure Demšar. 2024. Zero-shot reasoning: Personalized content generation without the cold start problem. arXiv [cs.AI] (Feb. 2024)

    2024

  15. [23]

    arXiv [cs.CL] (Sept

    Towards a unified view of preference learning for large Language Models: A survey. arXiv [cs.CL] (Sept. 2024)

    2024

  16. [24]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  17. [25]

    arXiv [cs.CL] (April 2024)

    Aligning LLM agents by learning latent preference from user edits. arXiv [cs.CL] (April 2024)

    2024

  18. [26]

    Peiling Jiang, Jude Rayan, Steven P Dow, and Haijun Xia. 2023. Graphologue: Exploring large language model responses with interactive diagrams. In Proceed- ings of the 36th annual ACM symposium on user interface software and technology . 1–20

    2023

  19. [27]

    Ruili Jiang, Kehai Chen, Xuefeng Bai, Zhixuan He, Juntao Li, Muyun Yang, Tiejun Zhao, Liqiang Nie, and Min Zhang. 2024. A survey on human preference learning for large language models. arXiv [cs.CL] (June 2024)

    2024

  20. [28]

    Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian Mcauley. 2023. Large language models as zero-shot conversational recommenders. InProceedings of the 32nd ACM International Conference on Informatio...

    2023

  21. [29]

    Ikram Karabila, Nossayba Darraz, Anas EL-Ansari, Nabil Alami, and Mostafa EL Mallahi. 2024. BERT-enhanced sentiment analysis for personalized e- commerce recommendations. Multimedia Tools and Applications 83, 19 (2024), 56463–56488

    2024

  22. [30]

    Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, and Prithviraj Ammanabrolu

  23. [31]

    Yoonsu Kim, Jueon Lee, Seoyoung Kim, Jaehyuk Park, and Juho Kim. 2024. Under- standing users’ dissatisfaction with chatgpt responses: Types, resolving tactics, and the effect of knowledge level. In Proceedings of the 29th International Confer- ence on Intelligent User Interfac...

    2024

  24. [32]

    Bart P Knijnenburg, Svetlin Bostandjiev, John O’Donovan, and Alfred Kobsa

  25. [33]

    Kai Konen, Sophie Jentzsch, Diaoulé Diallo, Peer Schütt, Oliver Bensch, Rox- anne El Baff, Dominik Opitz, and Tobias Hecking. 2024. Style vectors for steering generative large language model. arXiv [cs.CL] (Feb. 2024)

    2024

  26. [34]

    Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta, Maheswaran Sathiamoorthy, Lichan Hong, Ed Chi, and Derek Zhiyuan Cheng. 2023. Do LLMs understand user preferences? Evaluating LLMs on user rating prediction. arXiv [cs.IR] (May 2023)

    2023

  27. [35]

    Bruce W Lee, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Erik Miehling, Pierre Dognin, Manish Nagireddy, and Amit Dhurandhar. 2024. Programming refusal with Conditional Activation Steering. arXiv [cs.LG] (Sept. 2024)

    2024

  28. [36]

    Tae Soo Kim, Yoonjoo Lee, Minsuk Chang, and Juho Kim. 2023. Cells, generators, and lenses: Design framework for object-oriented interaction with large language models. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. 1–18

    2023

  29. [37]

    Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wat- tenberg. 2023. Inference-Time Intervention: Eliciting truthful answers from a language model. arXiv [cs.LG] (June 2023)

    2023

  30. [38]

    Xinyu Li, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized Human Feedback. arXiv [cs.CL] (Feb. 2024)

    2024

  31. [39]

    Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, et al. 2024. Personal llm agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459 (2024)

    2024 arXiv

  32. [40]

    Sheng Liu, Haotian Ye, Lei Xing, and James Zou. 2023. In-context vectors: Making in context learning more effective and controllable through latent space steering. arXiv [cs.LG] (Nov. 2023)

    2023

  33. [41]

    Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotiris Anagnostidis, Zhi Rui Tam, Keith Stevens, Abdullah Barhoum, Duc Nguyen, Oliver Stanley, Richárd Nagyfi, et al. 2024. Openassistant conversations-democratizing large language model alignment. Advances in Neural Informatio...

    2024

  34. [42]

    Zhenyi Lu, Wei Wei, Xiaoye Qu, Xianling Mao, Dangyang Chen, and Jixiong Chen

  35. [43]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing S...

    2020

  36. [44]

    Xiao Ma, Swaroop Mishra, Ariel Liu, Sophie Su, Jilin Chen, Chinmay Kulkarni, Heng-Tze Cheng, Quoc Le, and Ed Chi. 2023. Beyond ChatBots: ExploreLLM for structured thoughts and personalized model responses. arXiv [cs.HC] (Dec. 2023)

    2023

  37. [45]

    Sean M McNee, Shyong K Lam, Joseph A Konstan, and John Riedl. 2003. Interfaces for eliciting new user preferences in recommender systems. In International Conference on User Modeling . Springer, 178–187

    2003

  38. [46]

    Sheshera Mysore, Zhuoran Lu, Mengting Wan, Longqi Yang, Steve Menezes, Tina Baghaee, Emmanuel Barajas Gonzalez, Jennifer Neville, and Tara Safavi. 2023. Preprint, Google AR, 2024 Bo et al. PEARL: Personalizing large language model writing assistants with generation- calibrated...

    2023

  39. [47]

    Jonas Oppenlaender, Rhema Linder, and Johanna Silvennoinen. 2024. Prompting AI art: An investigation into the creative skill of prompt engineering.International Journal of Human–Computer Interaction (2024), 1–23

    2024

  40. [48]

    Ryan Louie, Andy Coenen, Cheng Zhi Huang, Michael Terry, and Carrie J Cai

  41. [49]

    In Proceedings of the 2020 CHI conference on human factors in computing systems

    Novice-AI music co-creation via AI-steering tools for deep generative models. In Proceedings of the 2020 CHI conference on human factors in computing systems. 1–13

    2020

  42. [50]

    Yingzhe Peng, Xiaoting Qin, Zhiyang Zhang, Jue Zhang, Qingwei Lin, Xu Yang, Dongmei Zhang, Saravan Rajmohan, and Qi Zhang. 2024. Navigating the un- known: A chat-based collaborative interface for personalized exploratory tasks. arXiv [cs.HC] (Oct. 2024)

    2024

  43. [51]

    arXiv [cs.CL] (Oct

    MIRACLE: Towards personalized dialogue generation with latent-space multiple personal attribute control. arXiv [cs.CL] (Oct. 2023)

    2023

  44. [52]

    Qianou Ma, Weirui Peng, Hua Shen, Kenneth Koedinger, and Tongshuang Wu

  45. [53]

    arXiv preprint arXiv:2409.08775 (2024)

    What you say= what you want? Teaching humans to articulate requirements for LLMs. arXiv preprint arXiv:2409.08775 (2024)

    2024 arXiv

  46. [54]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing . Association for Computational Linguistics. http://arxiv.org/abs/1908.10084

    2019 arXiv

  47. [55]

    Neil Rubens, Mehdi Elahi, Masashi Sugiyama, and Dain Kaplan. 2015. Active learning in recommender systems. Recommender systems handbook (2015), 809– 846

    2015

  48. [56]

    Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2023. LaMP: When large language models meet personalization. arXiv [cs.CL] (April 2023)

    2023

  49. [57]

    Alireza Salemi and Hamed Zamani. 2024. Comparing retrieval-augmentation and parameter-efficient fine-tuning for privacy-preserving personalization of large language models. arXiv [cs.CL] (Sept. 2024)

    2024

  50. [58]

    NA Osman, SA Mohd Noah, and M Darwich. 2019. Contextual sentiment based recommender system to provide recommendation in the electronic products domain. International Journal of Machine Learning and Computing 9, 4 (2019), 425–431

    2019

  51. [59]

    Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. 2023. Steering Llama 2 via Contrastive Activation Addi- tion. arXiv [cs.CL] (Dec. 2023)

    2023

  52. [60]

    Chengshuai Shi, Cong Shen, and Jing Yang. 2021. Federated multi-armed bandits with personalization. In International conference on artificial intelligence and statistics. PMLR, 2917–2925

    2021

  53. [61]

    Silviu Pitis, Ziang Xiao, Nicolas Le Roux, and Alessandro Sordoni. 2024. Improving context-aware preference modeling for language models. Advances in Neural Information Processing Systems 37 (2024), 70793–70827

    2024

  54. [62]

    Sriyash Poddar, Yanming Wan, Hamish Ivison, Abhishek Gupta, and Natasha Jaques. 2024. Personalizing reinforcement learning from Human Feedback with variational preference learning. arXiv [cs.LG] (Aug. 2024)

    2024

  55. [63]

    Alexandre Rame, Guillaume Couairon, Corentin Dancette, Jean-Baptiste Gaya, Mustafa Shukor, Laure Soulier, and Matthieu Cord. 2023. Rewarded soups: to- wards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. Advances in Neural Information Processi...

    2023

  56. [64]

    Sangho Suh, Meng Chen, Bryan Min, Toby Jia-Jun Li, and Haijun Xia. 2023. Luminate: Structured generation and exploration of design space with large language models for human-AI co-creation. arXiv [cs.HC] (Oct. 2023)

    2023

  57. [65]

    Zhaoxuan Tan and Meng Jiang. 2023. User modeling in the era of large language models: Current research and future directions. arXiv [cs.CL] (Dec. 2023)

    2023

  58. [66]

    Lev Tankelevitch, Viktor Kewenig, Auste Simkute, Ava Elizabeth Scott, Advait Sarkar, Abigail Sellen, and Sean Rintel. 2024. The metacognitive demands and opportunities of generative AI. In Proceedings of the CHI Conference on Human Factors in Computing Systems . 1–24

    2024

  59. [67]

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. Steering language models with activation engineering. arXiv [cs.CL] (Aug. 2023)

    2023

  60. [68]

    Scott Sanner, Krisztian Balog, Filip Radlinski, Ben Wedin, and Lucas Dixon

  61. [69]

    In Proceedings of the 17th ACM Conference on Recommender Systems, Vol

    Large language models are competitive near cold-start recommenders for language- and item-based preferences. In Proceedings of the 17th ACM Conference on Recommender Systems, Vol. 1. ACM, New York, NY, USA, 890–896

  62. [70]

    Omar Shaikh, Michelle Lam, Joey Hejna, Yijia Shao, Michael Bernstein, and Diyi Yang. 2024. Show, don’t tell: Aligning language models with demonstrated feedback. arXiv [cs.CL] (June 2024)

    2024

  63. [71]

    Rui Yang, Xiaoman Pan, Feng Luo, Shuang Qiu, Han Zhong, Dong Yu, and Jianshu Chen. 2024. Rewards-in-context: Multi-objective alignment of foundation models with dynamic preference adjustment. arXiv preprint arXiv:2402.10207 (2024)

    2024 arXiv

  64. [72]

    Ruizhe Shi, Yifang Chen, Yushi Hu, Alisa Liu, Hannaneh Hajishirzi, Noah A Smith, and Simon S Du. 2024. Decoding-time language model alignment with multiple objectives. arXiv [cs.LG] (June 2024)

    2024

  65. [73]

    Alessandro Stolfo, Vidhisha Balachandran, Safoora Yousefi, Eric Horvitz, and Be- smira Nushi. 2024. Improving instruction-following in language models through activation steering. arXiv [cs.CL] (Oct. 2024)

    2024

  66. [74]

    Nishant Subramani, Nivedita Suresh, and Matthew E Peters. 2022. Extracting latent steering vectors from pretrained language models. arXiv [cs.CL] (May 2022)

    2022

  67. [75]

    This is perfect, I like it!

    Yu Zhao, Alessio Devoto, Giwon Hong, Xiaotang Du, Aryo Pradipta Gema, Hongru Wang, Kam-Fai Wong, and Pasquale Minervini. 2024. Steering knowledge selection behaviours in LLMs via sae-based representation engineering. arXiv preprint arXiv:2410.15999 (2024). Steerable Chatbots: ...

    2024 arXiv

  68. [79]

    Dimitri von Rütte, Sotiris Anagnostidis, Gregor Bachmann, and Thomas Hofmann

  69. [80]

    arXiv [cs.CL] (Feb

    A language model’s guide through latent space. arXiv [cs.CL] (Feb. 2024)

    2024

  70. [81]

    Kaiwen Wang, Rahul Kidambi, Ryan Sullivan, Alekh Agarwal, Christoph Dann, Andrea Michi, Marco Gelmi, Yunxuan Li, Raghav Gupta, Avinava Dubey, et al

  71. [82]

    arXiv preprint arXiv:2407.15762 (2024)

    Conditional Language Policy: A General Framework for Steerable Multi- Objective Finetuning. arXiv preprint arXiv:2407.15762 (2024)

    2024 arXiv

  72. [83]

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences 68, 2 (2025), 121101

    2025

  73. [85]

    Xiwang Yang, Yang Guo, and Yong Liu. 2012. Bayesian-inference-based recom- mendation in online social networks.IEEE Transactions on Parallel and Distributed Systems 24, 4 (2012), 642–651

    2012

  74. [86]

    Dong Yi, Wang Zhilin, Makesh Narsimhan Sreedhar, Wu Xianchao, and Kuchaiev Oleksii. 2023. SteerLM: Attribute conditioned SFT as an (user-steerable) alterna- tive to RLHF. arXiv [cs.CL] (Oct. 2023)

    2023

  75. [87]

    JD Zamfirescu-Pereira, Richmond Y Wong, Bjoern Hartmann, and Qian Yang

  76. [88]

    In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems

    Why Johnny can’t prompt: how non-AI experts try (and fail) to design LLM prompts. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1–21

    2023

  77. [2007]

    InProceedings of the 16th international conference on World Wide Web

    Open user profiles for adaptive news systems: help or harm?. InProceedings of the 16th international conference on World Wide Web . 11–20

  78. [2012]

    In Proceedings of the sixth ACM conference on Recommender systems

    Inspectability and control in social recommenders. In Proceedings of the sixth ACM conference on Recommender systems . 43–50

  79. [2020]

    arXiv preprint arXiv:2010.12421 (2020)

    2020 arXiv

  80. [2023]

    arXiv [cs.CL] (Oct

    Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv [cs.CL] (Oct. 2023)

    2023

  81. [2024]

    arXiv preprint arXiv:2410.04070 (2024)

    Pad: Personalized alignment of llms at decoding-time. arXiv preprint arXiv:2410.04070 (2024)

    2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.