acceptodds
Withdrawn from ICLR 2027

Persona Dosing: Calibrated Activation Steering for Graded Trait Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

An activation-steering coefficient sets intervention strength, but requesting a particular degree of persona expression requires a behavioral scale. We study persona dosing: controlling a language model through a trait description and a requested mean intensity. PersonaDose specializes a shared, description-conditioned FLAS controller on persona responses, then calibrates its flow time against measured trait expression. Training responses are not paired with requested target intensities. Across Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B, PersonaDose raises core-trait expression at the Persona Vectors coherence floor of 75 by 33.2, 18.3, and 17.8 points over contrastive activation addition. Calibration-selected settings retain an expression advantage on held-out questions, although the coherence floor does not hold for every trait there. Across seven trained traits, calibrated requests yield mean targeting errors of 4.7 to 6.2 points over 14 to 22 calibration-reachable targets out of 28 per model. These results separate the behavioral range learned by a controller from the accuracy of requests within that range.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
activation steering, persona control, behavioral calibration, controllable text generation, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tCTUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper combines persona-based supervision, a shared activation controller, and post-training calibration to enable users to request different levels of trait expression based on a target mean score.

Critical strengths

Overall, I appreciate the distinction between what behaviors a model can express and how those behaviors can be requested on a behavioral score scale. This is a promising direction for interpretable research on LLM personality. Across three models and three core traits, the authors demonstrate stronger trait expression than CAA.

Critical weaknesses
  1. The introduction is not well written, which also makes the method difficult to understand. The third paragraph introduces PersonaDose directly, before sufficiently explaining the motivation and background. As a reader, I do not understand why this approach is needed, how its different stages and modules were motivated, or what each contributes. These aspects need a more detailed explanation.

  2. The evaluation lacks prompting and supervised fine-tuning (SFT), two standard intervention approaches. To my knowledge, AxBench found prompting to be a strong steering baseline. The proposed method itself uses cross-entropy training before calibration, so it remains unclear how much of its advantage comes from additional data and training. PersonaDose is trained on filtered persona responses, whereas the main direction-based baselines do not receive comparable response supervision.

  3. There are only ten test questions per trait. This provides limited diversity, and I am concerned that the evaluation does not offer sufficient coverage to assess the method reliably.

  4. The evaluation measures average scores across questions and responses, without guaranteeing that individual responses are close to the target. This definition allows substantially different output distributions to achieve equally good results.

Questions for authors

The paper’s main objective only requires the overall mean score to approach the requested value. This objective is clearly defined, but a low mean targeting error does not necessarily imply that responses exhibit a stable intermediate intensity. Mixing low- and high-intensity settings could also produce the same mean score.

I therefore hope the authors can explain what additional value continuous control provides over these simpler alternatives, such as more consistent performance across questions, fewer extreme responses, or lower control costs.

Reviewer_czomOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes Persona Dosing, a new way of steering LLM to exhibit certain traits (e.g., sycophantic). Previous methods generally work by giving contrastive input (e.g., different system prompts) and extract the difference between activations as steering vector. During inference, they transform activations with steering vector (e.g., shifting activation toward a steering vector). Existing methods generally suffer from degraded coherence when steered strongly. Also, existing methods are unable to control the degree of trait expression in an predictable way.

Persona Dosing transforms activations in a much more sophisticated way. It transforms the activation of the base model h(0) into h(T), by letting h(t) evolve by a differential equation:

h(t)/dt = v_theta (h(t), t, c),

from t = 0 -> T.

Here, v_theta is a (learned) function and c is a description of a trait. v_theta is trained in such a way that, after steering, the base model is more likely to produce outputs resembling those it produces when prompted to exhibit the trait described by c.

Experimental results show that it is able to steer the model while retaining coherence much better than previous methods. It is also able to meaningfully control the degree of model expression.

Critical strengths

(a) It proposes a method that steers a model much more robustly than previous methods while retaining coherence much better than previous methods. (b) The proposed method is also able to steer the model so that the response shows the degree of trait expression specified by the user.

  • Experimental results unambiguously show these strengths
Critical weaknesses
  • Writing could have been much better
Questions for authors

I like the approach and it solves the problem of coherence degradation by steering.

Do you have an intuitive explanation about why the method preserves coherence so well?

I did not like writing of the paper. It was incoherent and honestly painful to read. It is often the case that sentences in a paragraph are unconnected and do not have a logical flow with no emphasis on key ideas.

Reviewer_PRWMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of persona control for LLMs. The authors propose a model steering method persona dosing and evaluates it on 7 chosen personas. Results show clear improvements compared with several baselines across several model families, having questionable evaluation and baseline designs.

Critical strengths
  1. The empirical results reported in the paper show clear improvements, so the effectiveness of the method is valid.
  2. The authors have included a range of model families in their experiments, so the results are not dependent on one family of base models.
Critical weaknesses
  1. The method proposed in the paper seems to be highly similar to the FLAS framework. I doubt about the novelty for this work being no more than adopting the framework in persona control with an additionally trained controller.
  2. The paper misses several important related works in the literature about the topic of persona control or model steering for LLMs. The paper would benefit from engaging with these works: 2.1 HYPER STEER: Activation Steering at Scale with Hypernetworks 2.2 Bridging Mechanistic Interpretability and Prompt Engineering with Gradient Ascent for Interpretable Persona Control 2.3 ReFT-r1 in the AXBENCH paper: https://arxiv.org/pdf/2501.17148 2.4 Spherical Steering: Geometry-Aware Activation Rotation for Language Models
  3. The same LLM, GPT 4.1 mini, is used in both selecting training samples and evaluation, which seems to be a clear experimental misdesign.
  4. The main result on trait expression and coherence is based on a single model, GPT 4.1 mini, making the results highly model dependent on an underperforming model which has clear better options at this timepoint.
  5. The baselines chosen in this work seem to only include trivial fixes. There exist many stronger baselines, like HyperSteer and ReFT-r1, which are reported as comparable with the FLAS framework in the FLAS paper https://arxiv.org/pdf/2605.05892 are not included at all. Some very common persona steering methods like SAEs and prompt based methods are also not included as baselines.
Questions for authors
  1. Can the authors explain why and how the seven traits are chosen? I do not see an explicit explanation for this in the paper.
  2. Can the authors explain the differences between the method proposed in this work and the FLAS method we see in the literature?
  3. Can the authors explain why some commonly used baselines on persona steering are missed in this work, including SAE-based methods, prompt-based methods and FLAS-comparable methods like hyper steer and ReFT-r1?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.