acceptodds
Under review as a conference paper at ICLR 2027

TRAIT: selecTive and Robust Activation-space Intervention for Personality Trait Control

Abstract

Large language models are increasingly used to simulate human behavior, enabling social simulation experiments. In such experiments, researchers typically control the personality of these models through prompting, fine-tuning, or activation steering. However, existing methods still fall short in the selectivity and robustness of control, either perturbing unrelated dimensions or failing to sustain the preset level across contexts, thereby undermining experimental validity. To this end, we propose TRAIT, an automated framework for selective and robust personality control. On the extraction side, Direction Extraction from Deliberation Spans (DeliSpan) takes items from the psychometric scale of the target construct as seeds, has the target model synthesize its own decision scenarios, reason through them, and make decisions, and extracts direction vectors solely from the reasoning segments in which the model reveals its own preferences while weighing trade-offs. On the injection side, Dynamic Injection with Context-Adaptive Dosing (CtxDose) trains a dose network, supervised by the deviation between post-injection scale scores and preset scores, to predict the injection strength based on context. Evaluated on the IPIP-300 inventory within the psychometric framework of Serapio-Garc´ıa et al. (2025), TRAIT sets a new representation engineering state of the art in selectivity and robustness. These suggest personality can be controlled precisely and stably, providing a foundation for variable control in social simulation experiments.

Primary area:
foundation or frontier models, including LLMs
Keywords:
personality control, representation engineering, psychometric evaluation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_mYYFOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TRAIT, an activation-space framework for controlling LLM personality with an emphasis on selectivity and robustness. Its two main components are DeliSpan, which extracts steering directions from preference-bearing deliberation spans in model-generated decision scenarios, and CtxDose, which predicts context-dependent intervention strengths conditioned on a target personality score. The backbone model remains frozen. Experiments focus on Extraversion in Qwen3.8-27B using IPIP-300. Deliberation-span extraction achieves a larger selective operating amplitude than whole-trace extraction, while adaptive dosing reduces target-score RMSE from 0.225 with static dosing to 0.142. Prompting achieves lower target-score error but fails the paper’s selectivity criterion.

Critical strengths
  1. A meaningful focus on intervention selectivity. The paper evaluates changes in non-target personality dimensions alongside target-trait control. This is consequential for social simulation, where an intervention that changes several traits can confound the intended experiment. The comparison with prompting illustrates why target-score accuracy alone is insufficient.

  2. Complementary components supported by targeted ablations. DeliSpan and CtxDose address distinct sources of error: contamination in the extracted direction and context-dependent variation in intervention effects. Whole-trace extraction and static/projection-based dosing provide useful comparisons, with improvements in selective amplitude and target-score alignment supporting the proposed design.

Critical weaknesses
  1. The selectivity criterion relies on a debatable interpretation of human trait correlations. The metric treats human population correlations as the expected response of non-target traits to an intervention on Extraversion, and can penalize a non-target trait for remaining unchanged. Observational correlations do not by themselves establish these intervention effects. Furthermore, multiplying a correlation coefficient by a raw scale-score change requires assumptions about relative trait variances. Because this criterion determines the clean interval and much of the advantage over prompting, sensitivity to these assumptions is important for assessing the central claim.

  2. The validation is narrower than the claims about behavioral control and social simulation. Evaluation covers one model and one target trait, with robustness measured through aggregate questionnaire scores at four target levels. This establishes useful calibration evidence, but aggregate accuracy can conceal heterogeneous item-level errors and does not establish stable behavior in open-ended decisions or multi-turn interactions. I regard this as a limitation on the conclusions rather than evidence that the method is ineffective.

Questions for authors
  1. Clarification of the selectivity metric: What justifies interpreting human population correlations as expected causal co-movement under activation intervention? How do the method rankings change under an unadjusted non-target-drift criterion or a variance-adjusted regression formulation?

  2. Clarification of the evaluation protocol: Were the direction, injection layer, dose bounds, and all controller hyperparameters finalized without using the IPIP-300 results reported in Tables 1–2? Please specify which item sets were used for each selection stage.

  3. Clarification of robustness: Could you report the distribution of item-level errors or errors across held-out item subsets? This would help determine whether the aggregate RMSE improvement reflects consistent calibration or cancellation of positive and negative errors.

11 comments
AuthorsOpenReview ↗

Response to Reviewer mYYF [1/3]

Thank you for reviewing our work and recognizing its focus on intervention selectivity, the complementary designs of DeliSpan and CtxDose, and the corresponding ablation experiments.

We apologize for submitting our response so close to the end of the discussion period. We spent considerable time adding a comparison with a fine-tuning baseline, an evaluation in open-ended scenarios, and a joint evaluation of the system. As soon as the experiments concluded, we began compiling the results and preparing our point-by-point responses. Thank you for taking the time to read them before the discussion closes.

To W1 and Q1

We use human within-person trait covariation as a descriptive reference for assessing non-target drift. We measure changes in the model under intervention using the observed difference between steering and zero-dose conditions. Sₑ₅₉ still ranks first on composite selectivity when we account for the ratio of standard deviations or use raw drift directly.

In our social simulation task, isolating the target trait and preserving plausible relationships among traits are two aspects of the same control objective. We aim to manipulate Extraversion as an experimental variable while limiting how far the other traits depart from the covariation patterns observed in everyday human behavior. Here, covariation refers to correlations among personality traits within the same individual.

Repeated measurements by Ringwald et al. show that Extraversion's associations with the other four Big Five traits have the same signs at the between-person and within-person levels, with somewhat different magnitudes. The original manuscript therefore used between-person correlations to approximate within-person covariation. To make our analysis more rigorous and address your concern, we reviewed the literature further and recalculated selectivity using the within-person correlations reported by Ringwald et al. [1].

Multiplying the correlation coefficient directly by the change in Extraversion assumes equal within-person standard deviations for Extraversion and the non-target trait. To account for differences in variability across traits, we estimated the within-person standard deviations as σk,w=σk1−ICCk\sigma_{k,w}=\sigma_k\sqrt{1-\mathrm{ICC}_k} from the total standard deviations and intraclass correlations reported by Ringwald et al. [1]. To compute the reference drift, we also multiply by the ratio of the non-target trait's standard deviation to that of Extraversion.

Δ^k(α)=rk,wσk,wσE,wG(α).\widehat{\Delta}_k(\alpha)=r_{k,w}\frac{\sigma_{k,w}}{\sigma_{E,w}}G(\alpha).

Here, rk,wr_{k,w} is the within-person correlation between non-target trait kk and Extraversion, σk,w\sigma_{k,w} and σE,w\sigma_{E,w} are their respective within-person standard deviations, and G(α)G(\alpha) is the observed change in Extraversion score at dose α\alpha.

Under the correlation-adjusted criterion, the difference from the covariation reference contributes to the assessment even if the observed score for a non-target trait is unchanged. To address your concern directly, we also removed the correlation adjustment and reevaluated the five candidate vectors using raw drift in the non-target traits with the same sampling band. The table compares the three sets of results on the same dose grid and with the same sampling band.

VectorSelective operating amplitude with within-person correlation adjustment (points)Composite selectivity score with within-person correlation adjustmentSelective operating amplitude under the raw-drift criterion (points)Composite selectivity score under the raw-drift criterionComposite selectivity score with within-person correlation and standard deviation ratio adjustments
Sₑ₅₉1.233.563.767.3500.350
Sₑ₁₅.550.275.550.2750.275
Sₑ₄₃.650.290.600.2680.290
Tₑ₃₁.633.246.633.2460.246
Tₑ₃₅.750.348.400.1860.348

Sₑ₅₉ also ranks first among the five candidate vectors on selective operating amplitude under both the criterion using within-person correlations alone and the criterion using raw drift directly.

AuthorsOpenReview ↗

Response to Reviewer mYYF [2/3]

To W2

Using the open-ended scenarios and evaluation protocol from Persona Cartography, we extended our evaluation from questionnaires to free-form responses in everyday situations. The prompts describe concrete social and workplace situations in neutral language, such as “Should I go to a work happy hour I'm on the fence about?” The model gives advice or explains how it would respond in its own words. We then assess the personality tendencies expressed in each response using the Gemma judge that the study calibrated against human annotations. The evaluation includes 60 scenarios for each of the Big Five traits [2].

Vector selectivity. We selected the highest-scoring vector for selectivity from each extraction approach in the questionnaire evaluation, namely the deliberation-span vector Se59S_{e59} and the full reasoning-trace vector Te31T_{e31}. We compared their selectivity in open-ended scenarios using fixed-dose steering.

VectorSelective operating amplitude AseleA_{\mathrm{sele}} ↑Composite selectivity score SS ↑
Se59S_{e59}0.3420.175
Te31T_{e31}0.3000.143

The two vectors ranked in the same order as in the questionnaire evaluation. The results show that Se59S_{e59} retains a larger selective operating amplitude and a higher composite selectivity score in open-ended scenarios.

System robustness. We used the paper's inventory-level RMSE across four targets to compare target-control accuracy on 60 open-ended Extraversion scenarios, pairing the primary vector Se59S_{e59} with each of four dosing methods.

Dosing methodInventory-level RMSE across four targets ↓
Fixed dosing0.662
CtxDose0.570
MLP-based dosing0.832
Projection-based dosing0.920

The results show that in open-ended scenarios, CtxDose still brings inventory-level Extraversion scores closest to the prescribed targets. Its RMSE was the lowest of the four methods and 13.9% lower than with fixed dosing.

System selectivity. The paper's vector selectivity evaluation uses the steering dose α\alpha as the horizontal axis. The system selectivity evaluation uses the prescribed Extraversion target score TT. Since adaptive steering adjusts the dose to the context at a given target, we fixed the primary vector Se59S_{e59} and had all four dosing methods track the same 16 target scores. AseleA_{\mathrm{sele}} measures the span of observed Extraversion scores over which all four non-target traits satisfy the drift criterion.

Dosing methodSelective operating amplitude AseleA_{\mathrm{sele}} ↑Composite selectivity score SS ↑
Fixed dosing0.3420.175
CtxDose0.3500.177
MLP-based dosing0.2420.118
Projection-based dosing0.2750.138

The results show that with the primary vector Se59S_{e59}, CtxDose still has the largest selective operating amplitude and the highest composite selectivity score among the four dosing methods in open-ended scenarios.

AuthorsOpenReview ↗

Response to Reviewer mYYF [3/3]

To Q2

Table 1 reports a single IPIP-300 selectivity evaluation. We determined the five candidate vectors and their injection layers before evaluation and reported all five results. The selectivity and robustness evaluations are independent of each other. There is no stage in which the selectivity evaluation screens vectors for the robustness evaluation. The robustness experiment compares how closely each dosing method can steer Extraversion to the target with the same vector, items, and targets. The formal evaluation uses the 60 Extraversion items from IPIP-300. We used a separate set of 101 items to calibrate the doses and determine the controller hyperparameters.

DecisionData used and when the decision was madeDid the IPIP-300 sweep inform the decision?Basis for comparison
Five candidate vectors and their injection layersSelf-generated traces and 10 development items. The configurations were set before the sweep.No.We set all five configurations before seeing the IPIP-300 results and report all five in Table 1.
Vector used for the robustness evaluationAll dosing methods use Sₑ₅₉.The selectivity evaluation does not screen vectors for the robustness evaluation.We compare control error across dosing methods with the same vector, items, and targets.
Sweep doses and dose intervalsThe protocol prespecifies the sweep doses. We derive each vector's effective and selective intervals from the sweep results using the same criteria.The sweep doses were fixed. The reported intervals are evaluation outcomes.All candidates are evaluated at the prescribed doses, and all intervals are computed using the same criteria.
Dose calibration and controller hyperparameters for the robustness evaluationWe calibrated doses on an independent set of 101 calibration items, selected hyperparameters by cross-validation over 200 splits, and froze them before the formal robustness evaluation.No.The 101 calibration items are disjoint from IPIP-300. The formal evaluation uses the 60 Extraversion items to compute control error.
Base-model admission60 IPIP Extraversion items, 44 BFI items, and 20 PANAS items, for a total of 124 items. Each model underwent 250 administrations.Admission preceded the five-vector sweep and used the same 60 Extraversion items.The 60 items assess whether the base model meets the personality measurement requirements. We used the separate set of 101 items for controller training and dose calibration.

To Q3

To address your concern about context-specific differences, our robustness evaluation below also reports item-level RMSE. Let sˉi(T)\bar s_i(T) denote the mean score for item ii across four evaluation runs at target TT. We compute

RMSE⁡item(T)=160∑i=160(sˉi(T)−T)2.\operatorname{RMSE}_{\mathrm{item}}(T) =\sqrt{\frac{1}{60}\sum_{i=1}^{60}\bigl(\bar s_i(T)-T\bigr)^2}.

This metric measures how far individual item scores deviate from the target. Squaring each item's deviation ensures that scores both above and below the target contribute to the error.

We fixed the primary vector Se59S_{e59} and compared four dosing methods at the same four prescribed targets. We compute inventory-level RMSE from the errors in inventory mean scores at the four targets. The pooled value across all four targets uses all 240 item–target combinations.

Dosing methodInventory-level RMSE (points) ↓Pooled item-level RMSE across four targets (points) ↓
Fixed dosing0.1601.067
CtxDose0.1210.971
MLP-based dosing0.2321.002
Projection-based dosing0.1921.004

Compared with fixed dosing, CtxDose reduced inventory-level RMSE by about 24% and item-level RMSE by about 9%.

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

The manuscript targets a setting, where language models are used to simulate social behavior and where one wants to control the personality of the involved models reliably across different contexts and specifically enough to only target the traits one wants to influence. The article first establishes the personality type of the language model through standard psychometric inventories, in their case the Big Five model, and analyzes, whether that personality type is distinct enough to be influenced. Afterwards the framework called TRAIT distinguishes between two phases: extraction and injection. During the extraction phase, the model generates relevant scenarios for the trait to be influenced. Subsequently the model reasons through those scenarios with the reasoning either be retained in full or only the deliberations. This is done to match the target distribution. Based on retained part, the relevant layers are detected and their hidden states are aggregated. During injection, a light dose network is trained, where the dose strength depends on the context. The dose is only injected once after prefill. Afterwards the manuscript evaluates their approach and compares against three different baselines, while focusing on two metrics: selectivity (whether only the targeted trait is influenced) and robustness (whether the traits is consistently influenced across contexts). TRAIT shows in general better performance in both metrics than the baselines.

Critical strengths
  • clearly written
  • idea seams to be novel and the topic is a relevant to the research community by identifying, where personality traits sit inside a given model and how they can be influenced
  • reasonable evaluation with good ablation
Critical weaknesses

My biggest concern is the limited evaluation and applicability. Only a single model for a single trait are evaluated, which leaves the question whether the approach generalizes in either dimensions. From the tested three models, only one passed all requisites, which also seems to be somewhat limiting. The other larger concern is in regards to the personality type framework. Throughout the description of the TRAIT framework, it seems to be tied to the Big Five framework, even using one of its traits in Eq. 5, but that is never explicitly discussed, so again the question is whether the TRAIT approach generalizes to other personality type frameworks.

There is also no discussion of other relevant metrics such as how much compute is necessary to train the dosage network and what are the impacts on the end-to-end latency during inference.

  • minor:
    • I don't think it is a good idea to use SS as a symbol for the metric as well as for the deliberation-span vectors.
    • line 210 (potentially also before, for example line 127): I suggest to use the appropriate cite command for example \citet to avoid showing the name of the authors twice.
      • This comment applies in general to the manuscript, for example also for Appendix A.
    • references:
      • [Chen et al. 2025b] - missing place of publication
      • in general, the place of publication for arXiv preprints can only be derived from the URL
      • [Liu et al. 2026a] - I suggest to capitalize LLM to be consistent with the other references.
        • same for [Wu et al. 2025]
        • [Zou et al. 2023] I suggest to also capitalize AI.
    • C.2: might require some rewriting - some repetitions of facts/measured values
      • also somewhat applies to C.3
    • C.3: You might want to explain what S, T, R, and B0′ represent. This is only understandable from Figure 6.
Questions for authors

major questions:

  • How much compute resources were spent for training of dosage network? What is the impact of the framework on end-to-end latency?
  • How general is your approach in regards to the practical usage of language models? Only one out of three initial considered models passed the necessary requirements.
  • In the same vein as the previous question: Does your framework also generalize to other traits and potentially other personality type frameworks?

minor questions:

  • In the limitation, it is stated that prompting performs better for robustness, but I couldn't find the respective data: Where can I find that data?
  • How much is the framework tied to the psychometric scale and their levels? Examples always use five levels.
  • What does the index for the whole thinking trace/deliberation span mean?
  • Which LLM was used for judging the scenarios?
8 comments
AuthorsOpenReview ↗

Response to Reviewer hrsX [1/4]

Thank you for reviewing our work and recognizing the paper's clarity, the importance of the research question, and the value of the evaluations and ablations.

We apologize for submitting our response so close to the end of the discussion period. We spent considerable time adding a comparison with a fine-tuning baseline, an evaluation in open-ended scenarios, and a joint evaluation of the system. As soon as the experiments concluded, we began compiling the results and preparing our point-by-point responses. Thank you for taking the time to read them before the discussion closes.

To W1 and Q2 · Model admission

Larger, instruction-tuned models generally meet the admission requirements. Many models can pass, including Qwen3.5-9B and Qwen3-32B. We used the framework proposed by Serapio-García et al. to evaluate the reliability and validity of personality measurement in models. Of the 18 models evaluated in that study, 8 received favorable overall assessments, spanning multiple model families, including Flan-PaLM 62B and 540B, Llama 2-Chat 70B, Mixtral 8×7B Instruct, and GPT-4o. The study found that larger, instruction-tuned models typically exhibit a discernible personality structure [6].

We also evaluated the open-source models Qwen3.5-9B and Qwen3-32B against our admission criteria for Extraversion. Both passed all checks.

ModelIPIP Extraversion reliability αBFI Extraversion reliability αConvergent validity rDiscriminant validity gapAdmission outcome
Qwen3.8-27B (primary model in the paper)0.9610.9070.9170.638Pass
Qwen3.5-9B0.9440.8950.8930.481Pass
Qwen3-32B0.9530.8470.9020.509Pass
AuthorsOpenReview ↗

Response to Reviewer hrsX [2/4]

To W1, W2, Q2 and Q3 · Applicability across models, tendencies, and personality frameworks

A complete evaluation of generalization requires re-extracting steering vectors and calibrating doses for new models or tendencies, then evaluating selectivity and robustness separately. The time and GPU resources available during the discussion period did not allow us to complete this evaluation. To address your concern, we added a conformity steering experiment and drew on prior work to explain why TRAIT’s design supports strong generalization across models and tendencies and can accommodate any other construct with a scorable psychometric scale, extending well beyond the Big Five.

1. Conformity steering experiment

We used activation steering on Qwen2.5-7B-Instruct to target conformity. Based on the endpoints of the measured dose range, the aggregate conformity score increased by about 0.421 points per unit increase in dose, or about 16% relative to zero steering. In an open-ended multi-agent simulation, increasing the conformity steering dose from 0 to 0.8 raised the rate at which agents agreed with the visible majority from 40% to 90%. These results show that our approach can change a social behavior tendency beyond the Big Five in another model, with changes in both questionnaire scores and behavior in open-ended settings.

2. Theoretical analysis

  • Scope of representation engineering. Persona Vectors extracted steering vectors separately for Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct and found that activation steering could modify target behaviors such as sycophancy and hallucination. This provides a precedent for representation interventions across models and tendencies [2].
  • DeliSpan helps improve steering selectivity. Liu et al. found that vectors extracted directly from final responses can also change behavioral tendencies unrelated to the target. This makes it important to identify tokens that reflect how the model understands and weighs the target tendency and contain less information about unrelated tendencies. DeliSpan does this in three ways. First, the target model generates its own scenarios and reasoning traces, so the tokens used for extraction come from the model itself and reflect how it understands the target tendency. Second, items from psychometric scales serve as seeds for scenario synthesis, anchoring the extraction data in the behavioral scope of the target construct. Third, dilemmas elicit concrete trade-offs. DeliSpan extracts only the deliberation spans that express the target tendency and filters out background descriptions, restatements of facts, and irrelevant reasoning [3].
  • CtxDose helps improve robustness. In a refusal-control task, AdaSteer found that the same steering dose can have too little effect on some inputs and cause false refusals on others. This motivates adjusting the dose to the input [4]. Subsequent work also observed that trait expression scores in model outputs can initially rise and then fall as the dose increases [5]. There is no predetermined linear relationship between the projection of the hidden state onto the steering vector and the measured personality score. Applying AdaSteer's projection rule directly therefore makes it difficult to reliably steer personality scores to prescribed values across contexts. With CtxDose, we therefore explored using a network to predict the dose from the hidden state for the current context and the target score, training it on observed inventory scores. We tested MLP and linear models, and the linear model yielded lower target-control error in our experiments. One possible explanation is that near a given target, the required dose correction is approximately proportional to the context-specific score deviation. If the hidden state can predict this deviation linearly, a linear model suffices to estimate the required correction.
AuthorsOpenReview ↗

Response to Reviewer hrsX [3/4]

To W1, W2, Q2 and Q3 (continued)

3. TRAIT's design and experimental choices

  • TRAIT's general-purpose design. We designed TRAIT for general social simulation, with support for steering different psychological constructs. Any construct with a psychometric scale that can serve as seed data and a corresponding scoring rule can be incorporated into TRAIT's automated pipeline. Given the items and scoring rules, the system automatically synthesizes decision scenarios, extracts steering vectors, calibrates and applies doses, and evaluates the steering effect with the corresponding scale. Equation (5) uses the Extraversion score because Extraversion is the target in this paper. Under another personality framework, the target score, non-target dimensions, and the reference patterns of trait covariation can be updated based on the psychological evidence for the corresponding tendencies. The mechanisms for data synthesis, vector extraction, and dose control can be reused.
  • Model architecture. To demonstrate the generalizability of the TRAIT framework, we chose Qwen3.8-27B, a natively multimodal model with a hybrid attention architecture. Prior work found that representation interventions can be effective on text-only models such as Qwen2.5 and Llama 3.1. We therefore consider precise and robust control on a natively multimodal model with a different architecture as evidence that our approach generalizes well.
  • Target tendency. We chose the Big Five, the most widely studied framework in this area. Much related work uses steering performance within this framework to evaluate its methods. TRAIT's results within this framework demonstrate strong steering selectivity and robustness and are particularly relevant to progress in representation engineering and social simulation.

In summary, we addressed your concern about the scope of the method through the conformity steering experiment, prior work on representation interventions, and TRAIT’s design and experimental choices. The conformity experiment and prior work show that activation steering can modify tendencies beyond the Big Five across models. The theoretical analysis and the design of TRAIT explain how it can be adapted to other personality frameworks and why it has the potential to improve steering selectivity and robustness across models and traits.

To W3 and Q1 · Training cost and end-to-end latency

Cross-validation of the dose network over 200 splits took 54.35 minutes. At inference, CtxDose added 0.49 seconds (1.9%) of mean end-to-end latency relative to fixed dosing.

We first used the frozen base model on a GPU to sweep 16 dose levels across 101 calibration items, obtaining responses for 1,616 item–dose combinations. We ran cross-validation of the dose network on CPUs. Cross-validation over 200 splits took 54.35 minutes using 96 single-threaded processes in parallel.

We ran four paired evaluation passes on the official 60 Extraversion items, processing 240 requests serially per method. End-to-end timing ran from tokenization until the model generated its reasoning and returned the score for the selected response option. Adaptive methods also included real-time feature extraction by a resident instance of the same base model on a separate GPU and dose computation on a CPU. We started timing after model warm-up.

MethodMean latency (seconds)Median latency (seconds)95th-percentile latency (seconds)
No steering24.4124.6333.80
Fixed dosing25.9026.6335.17
CtxDose26.3926.8833.58
MLP-based dosing26.3826.7534.99

CtxDose increased mean end-to-end latency by 1.98 seconds (8.1%) relative to no steering and by 0.49 seconds (1.9%) relative to fixed dosing. Real-time feature extraction took 121 milliseconds on average, and dose computation took 0.644 milliseconds on average.

Reviewer_sHoeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies selective and robust personality control in language models through activation steering. It proposes extracting personality directions from preference-related segments of model-generated reasoning and adapting the intervention strength to the input context and desired trait level. The evaluation examines whether the method can control a target personality trait while limiting changes in other traits. The intended application is controlled social simulation, where both unintended effects and inconsistent interventions can undermine experimental validity.

Critical strengths
  • The paper addresses a meaningful problem for language-model-based social simulation. An intervention must do more than change the intended trait: it should also limit unintended changes and remain consistent across situations. Framing personality control around these requirements gives the work a clear methodological motivation.
  • The distinction between direction extraction and intervention calibration is useful. These components address different sources of failure, and separating them makes the proposed framework easier to analyze and potentially extend.
  • Extracting directions from reasoning segments that express preferences and trade-offs is a plausible approach to reducing irrelevant contextual information. The idea provides a concrete hypothesis about which parts of a generated response are useful for behavioral steering.
  • The use of psychometric instruments gives the control objective an explicit measurement framework.
  • The paper includes component comparisons and discusses limitations of its empirical scope. These analyses help identify where the method may be useful and where further validation is needed.
Critical weaknesses
  • The paper uses correlations between personality traits in human populations to specify acceptable changes under an intervention on a model. Observational covariation does not determine intervention effects, so preserving that structure cannot by itself demonstrate selective control. This assumption affects the interpretation of the central evaluation.
  • The statistical criterion for acceptable non-target changes is not sufficiently justified. Failure to detect a difference does not establish that the difference is practically negligible. A criterion based on sampling uncertainty may also become easier to satisfy when measurements are more variable, which is difficult to reconcile with the intended notion of precise control.
  • The connection between questionnaire responses and behavior. Controlling an aggregate personality score does not necessarily imply consistent decisions or interactions across situations. Since social simulation is the main motivation, the paper may need stronger evidence that the measured changes transfer to behavior outside the assessment format from my point of view.
  • The evidence for selectivity and robustness is not fully integrated. These properties are evaluated under different intervention settings, leaving uncertainty about whether the complete adaptive controller achieves both at once. This matters because input-dependent changes in intervention strength may alter non-target behavior even when the underlying direction performs well under fixed doses.
  • Appendix B.1 (p. 14) states that the final inventories are excluded from training and model selection. However, Appendix C.4 (p. 19) states that “the behavioral sweep doubles as a selection stage,” while Section 4.2 and Table 1 describe the reported sweep as an evaluation on an independent IPIP-300 test set. Were these separate sweeps on different item sets, or did the reported test results also inform vector selection? If the latter, the final evaluation would not be independent of candidate selection. Also I think that the baseline comparisons support a narrower conclusion than the broad claims of superiority.
Questions for authors
  • is the intended objective to isolate a personality dimension or to preserve human-like relationships between dimensions? How does the proposed selectivity criterion distinguish these objectives, and what justifies transferring between-person correlations to within-model interventions?
  • Were all choices concerning vectors, layers, and controller settings finalized before the reported test evaluations?
  • is there an evaluation of the complete adaptive controller that measures target control and non-target effects together in held-out behavioral contexts?
  • The reported valid region for Te31 appears inconsistent with the accompanying description. Appendix C.4 reports a decline in its target response at a higher dose, yet Table 5 retains that dose and subsequent doses within the valid region. Under the stated definition, a reversal beyond the uncertainty band should terminate this interval. Please clarify whether the reported decline meets that criterion and reconcile the table with the text.
10 comments
AuthorsOpenReview ↗

Response to Reviewer sHoe [1/5]

Thank you for reviewing our work and recognizing its focus on reliable personality control in social simulations, as well as the separation of steering vector extraction from dose calibration.

We apologize for submitting our response so close to the end of the discussion period. We spent considerable time adding a comparison with a fine-tuning baseline, an evaluation in open-ended scenarios, and a joint evaluation of the system. As soon as the experiments concluded, we began compiling the results and preparing our point-by-point responses. Thank you for taking the time to read them before the discussion closes.

To W1 and Q1

In our social simulation task, isolating the target trait and preserving plausible relationships among traits are two aspects of the same control objective. We aim to manipulate Extraversion as an experimental variable while limiting how far the other traits depart from the covariation patterns observed in everyday human behavior. Here, covariation refers to correlations among personality traits within the same individual.

Repeated measurements by Ringwald et al. show that Extraversion's associations with the other four Big Five traits have the same signs at the between-person and within-person levels, with somewhat different magnitudes. The original manuscript therefore used between-person correlations to approximate within-person covariation. To make our analysis more rigorous and address your concern, we reviewed the literature further and recalculated selectivity using the within-person correlations reported by Ringwald et al. [1].

To address your concern directly, we also removed the correlation adjustment and reevaluated the five candidate vectors using raw drift in the non-target traits with the same sampling band. The table reports selectivity under both criteria, one using within-person correlations and the other without correlation adjustment.

VectorSelective operating amplitude with within-person correlation adjustment (points)Composite selectivity score with within-person correlation adjustmentSelective operating amplitude under the raw-drift criterion (points)Composite selectivity score under the raw-drift criterion
Sₑ₅₉1.233.563.767.350
Sₑ₁₅.550.275.550.275
Sₑ₄₃.650.290.600.268
Tₑ₃₁.633.246.633.246
Tₑ₃₅.750.348.400.186

Under both criteria, Sₑ₅₉ ranks first among the five candidate vectors on both selective operating amplitude and composite selectivity score.

To W2

We use a sampling band to assess whether the observed non-target drift falls outside the range of sampling variation around the reference value. We estimate both the steering and zero-dose scores from a finite number of items, so the difference between the group means could vary if we used another set of comparable items. We set the width of the sampling band from the standard error of the difference between the two means to match the criterion to the measurement precision. The sampling band thus provides a statistically grounded way to assess whether drift relative to the reference value exceeds expected sampling variation.

To address your concern, we also evaluated fixed limits on score changes. We reevaluated the five candidate vectors at the original 15 doses with two criteria, limiting the absolute raw drift of each of the four non-target traits to 0.30 or 0.50 points. The fixed limits apply uniformly across vectors and doses and do not depend on sampling variation.

Table 1. Non-target drift limit of 0.30 points

Rank by SSVectorQualifying dose rangeSelective operating amplitude AseleA_{\mathrm{sele}} (points)Composite selectivity score SS
1Sₑ₅₉[−24, +8]0.4670.213
2Tₑ₃₅[−8, +4]0.4000.186
3Tₑ₃₁[−8, +4]0.4330.168
4Sₑ₄₃[−8, +4]0.3170.142
5Sₑ₁₅[−8, +8]0.2670.133

Table 2. Non-target drift limit of 0.50 points

Rank by SSVectorQualifying dose rangeSelective operating amplitude AseleA_{\mathrm{sele}} (points)Composite selectivity score SS
1Sₑ₅₉[−32, +16]0.7670.350
2Tₑ₃₅[−16, +8]0.7500.348
3Sₑ₄₃[−24, +8]0.6500.290
4Sₑ₁₅[−32, +8]0.5500.275
5Tₑ₃₁[−16, +8]0.6330.246

Under both fixed limits, Sₑ₅₉ ranks first among the five candidate vectors on both selective operating amplitude and composite selectivity score.

AuthorsOpenReview ↗

Response to Reviewer sHoe [2/5]

To W3

Using the open-ended scenarios and evaluation protocol from Persona Cartography, we extended our evaluation from questionnaires to free-form responses in everyday situations. The prompts describe concrete social and workplace situations in neutral language, such as “Should I go to a work happy hour I'm on the fence about?” The model gives advice or explains how it would respond in its own words. We then assess the personality tendencies expressed in each response using the Gemma judge that the study calibrated against human annotations. The evaluation includes 60 scenarios for each of the Big Five traits [2].

Vector selectivity. From the questionnaire evaluation, we selected the deliberation-span vector Se59S_{e59} and the full reasoning-trace vector Te31T_{e31}, the vector with the highest selectivity score from each extraction approach. We compared their selectivity in open-ended scenarios using fixed-dose steering.

VectorSelective operating amplitude AseleA_{\mathrm{sele}} ↑Composite selectivity score SS ↑
Se59S_{e59}0.3420.175
Te31T_{e31}0.3000.143

The two vectors ranked in the same order as in the questionnaire evaluation. The results show that Se59S_{e59} retains a larger selective operating amplitude and a higher composite selectivity score in open-ended scenarios.

System robustness. We used the paper's inventory-level RMSE across four targets to compare target-control accuracy on 60 open-ended Extraversion scenarios, pairing the primary vector Se59S_{e59} with each of four dosing methods.

Dosing methodInventory-level RMSE across four targets ↓
Fixed dosing0.662
CtxDose0.570
MLP-based dosing0.832
Projection-based dosing0.920

The results show that CtxDose still brings inventory-level Extraversion scores closest to the prescribed targets in open-ended scenarios. It had the lowest RMSE of the four methods, 13.9% below fixed dosing.

System selectivity. The paper's vector selectivity evaluation uses the steering dose α\alpha as the horizontal axis. The system selectivity evaluation uses the prescribed Extraversion target score TT. Since adaptive steering adjusts the dose to the context at a given target, we fixed the primary vector Se59S_{e59} and had all four dosing methods track the same 16 target scores. AseleA_{\mathrm{sele}} measures the span of observed Extraversion scores over which all four non-target traits satisfy the drift criterion.

Dosing methodSelective operating amplitude AseleA_{\mathrm{sele}} ↑Composite selectivity score SS ↑
Fixed dosing0.3420.175
CtxDose0.3500.177
MLP-based dosing0.2420.118
Projection-based dosing0.2750.138

The results show that with the primary vector Se59S_{e59}, CtxDose still has the largest selective operating amplitude and the highest composite selectivity score among the four dosing methods in open-ended scenarios.

AuthorsOpenReview ↗

Response to Reviewer sHoe [3/5]

To W4 and Q3

1. Robustness of the final controller

The robustness results in the paper are for the full controller. We fixed the primary vector Se59S_{e59} from the paper and paired it with each of four dosing methods to measure Extraversion control error at the same four targets.

Dosing methodRMSE across four targets (points) ↓
Fixed dosing0.160
CtxDose0.121
MLP-based dosing0.232
Projection-based dosing0.192

When paired with Se59S_{e59}, CtxDose achieved the lowest target-control error, about 24% below fixed dosing.

2. Selectivity of the final controller

The paper evaluates vector selectivity by sweeping the steering dose α\alpha. For the full controller, users set a target Extraversion score TT and each method determines its dose, with CtxDose also adjusting the dose for each item. We therefore changed the horizontal axis of the selectivity evaluation from steering dose to prescribed target score and evaluated both vectors with all four dosing methods at the same set of prescribed target scores.

For each combination, we identified the contiguous sequence of target values closest to the prescribed reference target at which all four non-target traits met the drift criterion. AseleA_{\mathrm{sele}} is the span of observed Extraversion scores over those target values, and SS is the composite selectivity score. For example, with Se59S_{e59} paired with CtxDose, the qualifying targets ranged from 1.0 to 3.7 and observed Extraversion scores ranged from 2.400 to 3.667, giving a selective operating amplitude of 1.267 points.

Selective operating amplitude (points)

Dosing methodSe59S_{e59}Te31T_{e31}
Fixed dosing1.0170.550
CtxDose1.2670.883
MLP-based dosing1.0500.667
Projection-based dosing0.9830.750

Composite selectivity score

Dosing methodSe59S_{e59}Te31T_{e31}
Fixed dosing0.3530.200
CtxDose0.4470.292
MLP-based dosing0.3810.222
Projection-based dosing0.3520.274

For each dosing method, Se59S_{e59} scored higher than Te31T_{e31} on both selectivity metrics. With either vector, CtxDose scored highest on both metrics among the four methods. Overall, Se59S_{e59} paired with CtxDose achieved the lowest target error in the four-method comparison with the primary vector and the highest selectivity among all eight combinations. The two system evaluation tables in the response to W3 report the full controller's robustness and selectivity in open-ended scenarios.

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

The paper introduces TRAIT, an activation-space intervention framework for selective and robust personality control in large language models. DeliSpan extracts steering directions from preference and trade-off spans in model-generated reasoning, while CtxDose predicts context-dependent intervention strengths using a controller trained on offline dose–response curves. Experiments on Extraversion control in Qwen3.8-27B using IPIP-300 show a wider selective control range under the paper’s criterion and lower target-score error than fixed-dose and projection-based controls.

Critical strengths
  1. Practically relevant problem. The paper addresses both target accuracy and off-target changes, tackling an important obstacle to reliable variable control in LLM-based social simulations.

  2. Clear methodological design. DeliSpan focuses on preference-bearing reasoning spans, while CtxDose adjusts intervention strength to context. Both components have clear motivations.

Critical weaknesses
  1. Limited evaluation scope. Full-method evaluation covers only one model and Extraversion, mainly through single-turn questionnaires. Generalization to other models, traits, and multi-turn interactions remains unverified.

  2. *Limited joint validation. Inventory averages may hide context-specific errors, and separate fixed-dose and adaptive-dose tests do not establish that the final controller is both accurate and selective.

  3. Limited comparisons and ablations. Fine-tuning baselines are absent, several prior methods are evaluated through adaptations, and Figure 3 changes both extraction spans and layers, making it difficult to isolate the effect of span selection.

Questions for authors
  1. Could you evaluate the complete method on additional models, target traits, or multi-turn interactions to establish its generalizability?

  2. Could you report context-level target errors and off-target drift for the final CtxDose controller to establish accuracy and selectivity jointly?

  3. Could you add a fine-tuning baseline and compare span-based versus full-reasoning extraction at matched extraction layers and injection settings to isolate the contribution of span selection?

9 comments
AuthorsOpenReview ↗

Response to Reviewer ciuY [1/5]

Thank you for reviewing our work and recognizing its relevance to reliable variable control in social simulations and the motivation behind DeliSpan and CtxDose.

We apologize for submitting our response so close to the end of the discussion period. We spent considerable time adding a comparison with a fine-tuning baseline, an evaluation in open-ended scenarios, and a joint evaluation of the system. As soon as the experiments concluded, we began compiling the results and preparing our point-by-point responses. Thank you for taking the time to read them in the remaining discussion period.

To W1 and Q1

We extended the evaluation of the full method for Extraversion control from questionnaires to free-form responses in open-ended scenarios. TRAIT retained its advantages in both selectivity and robustness. To clarify where the method can apply across models and tendencies, we draw on our conformity experiment, prior work on representation interventions, and TRAIT's design.

1. Evaluation in open-ended scenarios

Using the open-ended scenarios and evaluation protocol from Persona Cartography, we extended our evaluation from questionnaires to free-form responses in everyday situations. The prompts describe concrete social and workplace situations in neutral language, for example, “Should I go to a work happy hour I'm on the fence about?” The model gives advice or explains how it would respond in its own words. We then assess the personality tendencies expressed in each response using the Gemma judge that the study calibrated against human annotations. The evaluation includes 60 scenarios for each of the Big Five traits [1].

Vector selectivity. From the questionnaire evaluation, we selected the highest-scoring vector for selectivity from each of two extraction approaches, the deliberation-span vector Se59S_{e59} and the full reasoning-trace vector Te31T_{e31}. We compared their selectivity in open-ended scenarios using fixed-dose steering.

VectorSelective operating amplitude AseleA_{\mathrm{sele}} ↑Composite selectivity score SS ↑
Se59S_{e59}0.3420.175
Te31T_{e31}0.3000.143

The two vectors ranked in the same order as in the questionnaire evaluation. In open-ended scenarios, Se59S_{e59} still had a larger selective operating amplitude and a higher composite selectivity score.

System robustness. We used the paper's inventory-level RMSE across four targets to compare target-control accuracy on 60 open-ended Extraversion scenarios, pairing the primary vector Se59S_{e59} with each of four dosing methods.

Dosing methodInventory-level RMSE across four targets ↓
Fixed dosing0.662
CtxDose0.570
MLP-based dosing0.832
Projection-based dosing0.920

In open-ended scenarios, CtxDose still brought the inventory-level Extraversion scores closest to the prescribed targets. Its RMSE was 13.9% lower than with fixed dosing.

System selectivity. The paper's vector selectivity evaluation uses the steering dose α\alpha as the horizontal axis. Our system selectivity evaluation uses the prescribed Extraversion target score TT. Since adaptive steering adjusts the dose to the context at a given target, we fixed the primary vector Se59S_{e59} and evaluated the four dosing methods on the same 16 target scores. AseleA_{\mathrm{sele}} measures the span of observed Extraversion scores over which all four non-target traits satisfy the drift criterion.

Dosing methodSelective operating amplitude AseleA_{\mathrm{sele}} ↑Composite selectivity score SS ↑
Fixed dosing0.3420.175
CtxDose0.3500.177
MLP-based dosing0.2420.118
Projection-based dosing0.2750.138

In open-ended scenarios, the primary vector Se59S_{e59} paired with CtxDose still achieved the largest selective operating amplitude and the highest composite selectivity score.

AuthorsOpenReview ↗

Response to Reviewer ciuY [2/5]

To W1 and Q1 (continued)

2. Applicability across models and tendencies

Conformity steering experiment. We used activation steering on Qwen2.5-7B-Instruct to target conformity. Comparing the endpoints of the measured interval, the aggregate conformity score on a structured questionnaire increased by about 0.421 points per unit of dose, an increase equivalent to about 16% of the zero-dose score. In an open-ended multi-agent simulation, increasing the conformity steering dose from 0 to 0.8 increased the rate at which agents' positions aligned with the visible majority from 40% to 90%. These results show that activation steering can adjust a tendency beyond the Big Five in another model, with changes in both questionnaire scores and simulated behavior.

Theoretical analysis

  • Scope of representation engineering. Persona Vectors extracted steering vectors separately for Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct and found that activation steering could modify target behaviors such as sycophancy and hallucination. This provides a precedent for representation interventions across models and tendencies [2].
  • DeliSpan helps improve steering selectivity. Liu et al. found that vectors extracted directly from final responses can also change behavioral tendencies unrelated to the target [3]. This makes it important to identify tokens that reflect how the model understands and weighs the target tendency and contain less information about unrelated tendencies. DeliSpan does this in three ways. First, the target model generates its own scenarios and reasoning traces, so the tokens used for extraction come from the model itself and reflect how it understands the target tendency. Second, items from psychometric scales serve as seeds for scenario synthesis, anchoring the extraction data in the behavioral scope of the target construct. Third, dilemmas elicit concrete trade-offs. DeliSpan extracts only the spans that express the target tendency and filters out background descriptions, restatements of facts, and irrelevant reasoning.
  • CtxDose helps improve robustness. In a refusal-control task, AdaSteer found that the same steering dose can have too little effect on some inputs and cause false refusals on others. This motivates adjusting the dose to the input [4]. Subsequent work also observed that trait expression scores in model outputs can initially rise and then fall as the dose increases [5]. There is no predetermined linear relationship between the projection of the hidden state onto the steering vector and the measured personality score. Applying AdaSteer's projection rule directly therefore makes it difficult to reliably steer personality scores to prescribed values across contexts. CtxDose first uses the average dose–response curve to set a base dose. It then predicts a dose correction from the current context's hidden state and the target score, using observed inventory scores as supervision. We tested MLP and linear models, and the linear model yielded lower target-control error in our experiments. One possible explanation is that near a given target, the required dose correction is approximately proportional to the context-specific score deviation. If the hidden state can predict this deviation linearly, a linear model suffices to estimate the required correction.
AuthorsOpenReview ↗

Response to Reviewer ciuY [3/5]

To W1 and Q1 (continued)

TRAIT's design and experimental choices

  • TRAIT's general-purpose design. We designed TRAIT for general social simulation, with support for steering different psychological constructs. Any construct with a psychometric scale that can serve as seed data and a corresponding scoring rule can be incorporated into TRAIT's automated pipeline. Given the items and scoring rules, the system automatically synthesizes decision scenarios, extracts steering vectors, calibrates and applies doses, and evaluates the steering effect with the corresponding scale. Equation (5) uses the Extraversion score because Extraversion is the target in this paper. Under another personality framework, the target score, non-target dimensions, and the reference patterns of trait covariation can be updated based on the psychological evidence for the corresponding tendencies. The mechanisms for data synthesis, vector extraction, and dose control can be reused.
  • Model architecture. To examine TRAIT's generalizability, we chose Qwen3.8-27B, a natively multimodal model with a hybrid attention architecture. Prior work found that representation interventions can be effective on text-only models such as Qwen2.5 and Llama 3.1. Precise and robust Extraversion control on a natively multimodal model with a different architecture extends the evidence for these methods to another architecture.
  • Target tendency. We chose the Big Five, the most widely studied framework in this area. Much related work uses steering performance within this framework to evaluate its methods. TRAIT's results within the framework show that it can steer Extraversion selectively and robustly. This is a problem that representation engineering must address to support social simulation.

The open-ended evaluation directly addresses your concern that the evaluation relies primarily on questionnaires. We extended the evaluation of the full method to free-form responses in everyday situations, and the results show that TRAIT's advantages in selectivity and robustness for Extraversion control extend to open-ended scenarios.

We addressed your concern about applicability across models and tendencies through the conformity steering experiment, theoretical analysis, and TRAIT's design and experimental choices. The conformity results and prior work show that activation steering can adjust tendencies beyond the Big Five across models. The theoretical analysis and the design of TRAIT explain how it can be adapted to other personality frameworks and why it has the potential to improve steering selectivity and robustness across models and traits.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the Area Chair

We thank you and the reviewers for taking the time to review our paper. The reviewers recognized the importance of steering target tendencies selectively and robustly in social simulation. They also recognized the effectiveness of the design choices behind DeliSpan and CtxDose.

We apologize for submitting our response so close to the end of the discussion period. We spent considerable time adding a comparison with a fine-tuning baseline, an evaluation in open-ended scenarios, and a joint evaluation of the system. As soon as the experiments concluded, we began compiling the results and preparing our point-by-point responses. Thank you all for taking the time to read them before the discussion closes.

The following five points address the reviewers' questions and support the paper's core conclusions.

  1. TRAIT achieved better selectivity and robustness than the fine-tuning baseline. With the same base model and the same 1,026 training traces, TRAIT adjusted Extraversion over a range about 7.4 times as wide as the baseline while keeping the other traits stable. The error between the mean questionnaire score and the prescribed target was about 50% lower.
  2. TRAIT ranked first in selectivity and robustness in the evaluation of the full system, including vector extraction and dose application. Compared with fixed dosing, CtxDose had about 24% lower target-score error and about 25% higher scores on both selectivity metrics.
  3. TRAIT retained its selectivity and robustness advantages in open-ended scenarios relevant to social simulation. The vector extracted with DeliSpan continued to show higher selectivity, and CtxDose achieved the lowest target-score error and the highest system selectivity among the four dosing methods.
  4. Psychometric scales provide TRAIT's data and anchor its constructs. TRAIT uses individual scale items as seeds to synthesize data for steering-vector extraction. Any psychological construct with a psychometric scale that can provide seed data and a corresponding scoring rule can be incorporated into TRAIT's automated pipeline.
  5. Applicability across models and traits. We addressed applicability across models and traits through the conformity steering experiment, theoretical analysis, and TRAIT's design and experimental choices. The conformity experiment and prior studies support the feasibility of steering different tendencies across models. Our theoretical analysis and design show that TRAIT can be adapted to other personality frameworks and has the potential to improve steering selectivity and robustness across models and traits.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.