acceptodds
Under review as a conference paper at ICLR 2027

Persona Cartography: Charting Language Model Personality Traits in Weight Space

10 views

Abstract

Large language models exhibit recurring behavioural patterns — personas — that shape generalisation and safety, but we lack reliable tools for decomposing, measuring, and controlling them. Our central insight is to treat personas as positions in a space of behavioural traits, using the OCEAN framework to describe model personas in terms of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. We train low-rank adapters to amplify or suppress individual traits, and evaluate their effects using an LLM-judge calibrated against independent crowdsourced human raters, trait-specific multiple-choice benchmarks, and standard capability evaluations. Across six models from three families (4B-32B), we find that each adapter moves its target trait largely monotonically with scale, combines approximately additively with other adapters to construct mixed personas, and preserves performance on capability benchmarks at moderate scales. We further show that the induced trait axes affect safety-relevant behaviour in downstream evaluations: for example, moving along the neuroticism axis modulates frustration on impossible tasks, and composing adapters shifts the trade-off between harmful compliance and over-refusal. We also take a first, exploratory step beyond human trait taxonomies: an unsupervised psychometric pipeline that recovers four candidate behavioural factors (tone, initiative, didacticism, epistemic caution) from synthetic model rollouts, three of which survive controlling for conversational scenario. Together, these results frame persona control as learning, scaling, and composing traits in weight space.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Alignment, personas

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_DXe3OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates whether the personality-like behaviors exhibited by Large Language Models (LLMs) can be evaluated, controlled, and combined within the weight space. Grounded in the classic OCEAN "Big Five" personality framework, the authors first define the behavioral characteristics associated with each personality trait. They then employ DPO and SFT techniques in a two-stage training process to develop LoRA adapters capable of either enhancing or suppressing specific personality dimensions. Subsequently, the authors conduct systematic experiments to evaluate whether these LoRAs can achieve personality control—both individually and in combination—thereby enabling the steering of the model toward specific personality profiles (mixed personality control). They also verify the method's generalizability across different models and teacher configurations. Furthermore, the authors explore downstream applications through "Frustration" and "WildJailbreak" experiments, demonstrating that personality modulation influences safety-related behaviors and holds practical value. Finally, recognizing that OCEAN is a classification framework originally designed for humans, the authors use unsupervised factor analysis on model interaction trajectories to extract latent, "AI-native" behavioral factors. This leads to the formulation of the TIDE framework and a preliminary exploration of the controllability of these factors, revealing significant differences in control efficacy based on the specific factor and the direction of intervention.

Critical strengths
  1. The research addresses a problem of practical significance and demonstrates potential value for downstream applications: As LLM application scenarios expand, understanding and modulating the behavioral tendencies of the models themselves is of real-world importance. The intervention method proposed in this paper offers a means to steer model behavior toward desired personality traits. Furthermore, the study goes beyond merely verifying personality control; through downstream experiments such as "Frustration" and "WildJailbreak," it demonstrates that adjusting model personality influences safety-relevant downstream behaviors, thereby highlighting the practical value and potential of this research direction.
  2. The paper features a cohesive narrative and comprehensive experiments, establishing a clear and logical chain of argumentation: The authors begin with the classic OCEAN personality theory to define and evaluate model personality traits. They then develop and implement a personality intervention method and conclude by discussing various control strategies (including individual, scaled, and combined control effects) alongside a thorough validation of robustness and generalizability. The overall presentation is highly complete.
  3. The paper goes beyond predefined human personality frameworks to explore latent behavioral factors in LLMs: The authors recognize and actively investigate the possibility that the concept of "persona" derived from human studies may not be directly applicable to LLMs, designing experiments to probe this distinction. Although this exploration is in its early stages, it offers crucial insights for the research community, particularly for future studies focused on LLM personas.·
Critical weaknesses
  1. The novelty of the core intervention method is somewhat limited, although the paper offers distinct empirical and exploratory contributions. By employing LoRA to intervene in LLM personas, the paper does not propose a completely new paradigm for personality control but largely follows the technical trajectory of existing research. The authors acknowledge these connections to prior work, which I appreciate; however, this transparency does not eliminate the relatively incremental nature of the core intervention method. I believe the paper's true value lies in: 1. a systematic and comprehensive experimental design that validates the effectiveness, adjustability, composability, and robustness of personality intervention, together with downstream behavioral validation; 2. methodological transparency and experimental rigor; and 3. proactive reflection on existing research paradigms and the exploration of latent behavioral factors in LLMs via TIDE. Therefore, the paper could better highlight these distinct empirical and exploratory contributions, while recognizing that the novelty of the central intervention technique itself remains somewhat limited.

  2. The scope of comprehensive validation for core experiments remains limited; specifically, the cross-model generalizability of TIDE has not been fully verified. Although the authors candidly acknowledge in the limitations section that the comprehensive evaluation framework focused primarily on Llama-3.1-8B-Instruct, this constraint still poses potential risks regarding generalizability. While the authors trained personality LoRAs on six LLMs and verified the method's applicability through select metrics and combination experiments, the full evaluation suite and most in-depth analyses remain centered on Llama-3.1-8B-Instruct. This issue is particularly pronounced for TIDE: the process of identifying key factors was based on interaction trajectories generated by Llama-3.1-8B-Instruct. Despite conducting a partial cross-model consistency check using Qwen2.5-7B-Instruct, the authors have not independently replicated the full factor discovery and control workflow on other models. Consequently, the current results may still be influenced by model-specific behavioral patterns, offering relatively limited support for TIDE's generalizability.

  3. Evaluations relying on an LLM Judge entail uncertainties regarding reliability. The paper employs LLM judges for automated evaluation across several experiments. While the authors performed independent human calibration for the primary personality-scoring Judge, the Judges used for certain downstream safety-related evaluations did not undergo the same level of validation. These automated judges may still introduce additional judgment noise or systematic bias; however, these experiments provide crucial evidence supporting the downstream practical value of personality interventions, and additional validation would help establish the reliability of these downstream findings.

Questions for authors
  1. The authors could further clarify the paper's contribution positioning in the Introduction, Contributions, and Related Work sections. Although the paper already acknowledges its relationship to prior studies, it would be helpful to more explicitly distinguish the inherited technical approaches from the paper's distinct contributions, particularly the systematic experimental validation of LoRA scaling and composition, downstream behavioral effects, methodological transparency, and the exploratory TIDE findings.

  2. The overall presentation could be further polished, especially for some information-dense figures (e.g., Figure 4、5). Improvements to color contrast, label readability, legend clarity, and panel layout would make the experimental results easier to interpret and improve the overall reading experience.

  3. The authors could consider supplementing the downstream evaluations (e.g., Frustration and WildJailbreak) with additional human–LLM agreement evidence. While the main personality-scoring judges have been calibrated against human raters, comparable validation is limited for the downstream judges. Clarifying whether any existing human validation is available, or providing a small-scale human agreement analysis if feasible, would help strengthen confidence in these downstream findings.

2 comments

Thank you very much for the thorough review. Here are responses to the various comments and suggestions:

  1. We agree with the critique that we could state more clearly how our novel contributions build upon existing work. We will clarify which of our contributions are novel and which are inherited from prior work, and will do this in the suggested sections.
  2. We agree that the figures can be very information dense, and so we will endeavour to increase readability and implement the suggested improvements to the figures.
  3. We appreciate that it would be better to have human LLM-judge calibration for all judge use cases, and that some noise or bias can be introduced by not doing so. For the downstream behaviour judges we replicated the approaches from pre-existing experiments, and opted to do so in order to make our results comparable to prior work - calibrating these judges would have made this comparison harder. Given the choice between the two, we opted to use the judges unchanged. Our benign-compliance judge was the only novel downstream effect judge which was not taken from prior work and was not human calibrated.
  4. We agree with the critiques of the TIDE framework. We would like to emphasize that this section is an exploratory proof of concept, which we hope will spur further work. We will make that clearer in the main body of the paper.
Reviewer_DXe3OpenReview ↗

response to author

I appreciate the author’s clarifications and commitment to making revisions. I will maintain my current score, as I consider this a solid piece of work that deserves to be accepted by ICLR. I look forward to future research that moves beyond human-centric frameworks to explore the inherent "personality" of the models themselves.

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

This paper adapts principles from human personality psychology—specifically the Big Five / OCEAN framework—to characterize and steer large language model (LLM) personas within weight space. The authors propose training trait-specific LoRA adapters using a constitution-guided distillation pipeline (adapted from Maiya et al.) to achieve continuous, controllable persona modulation via weight manipulation. The paper presents three primary contributions: 1) controllable steering of LLM persona traits in weight space using trait-specific LoRA adapters, 2) demonstration of downstream utility and safety implications resulting from persona modulation, and 3) an unsupervised factor analysis framework designed to discover latent persona dimensions beyond traditional OCEAN bases. The methodology is validated empirically across multiple model families and parameter scales, demonstrating both the feasibility of weight-space persona steering and the emergence of latent behavioral factors.

Critical strengths
  1. Compelling Framing and Grounded Behavioral Modeling: Bridging human personality psychology (the OCEAN framework) with weight-space manipulation offers a novel and principled perspective on LLM alignment, steerability, and persona characterization. Understanding behavioral traits at the parameter level is fundamentally important for both interpretability and controllable generation.

  2. Solid Technical Execution and Broad Empirical Validation: The methodology is technically sound and thoroughly evaluated across multiple model families and scales. Demonstrating both supervised steering via LoRA adapters and unsupervised latent persona discovery provides meaningful empirical depth that substantially supports the paper's claims

Critical weaknesses
  1. Informal Methodological Formulation: The core methodology in Section 2.1 relies heavily on high-level textual descriptions rather than rigorous mathematical formalisms. Specifically, the definitions and operations surrounding the weight space construction, linear persona steering/interpolation, and the factor-analytic decomposition of persona vectors remain ambiguous. A self-contained, formal mathematical framing is necessary to ensure reproducibility and theoretical clarity.

  2. Draft-Quality Exposition and Structure: The overall presentation feels unpolished in several sections. The narrative flow is disrupted by numerous single-sentence paragraphs, and several technical terms/metrics are introduced without adequate contextualization or definitions. Addressing these structural and drafting shortcomings is critical for the work to meet conference publication standards.

Questions for authors

***Clarifications Impacting the Assessment:

  1. LLM-as-a-Judge Robustness: The supervised labels rely on an LLM-as-a-judge setup. How sensitive are the downstream personality measurements and distillation results to the choice of the judge model? Have you evaluated consistency across different evaluator models or human annotations?

  2. Mathematical and Operational Definitions:

--- In Line 237, could you formally clarify what is meant by "...interpolation between instruct-tuned and base models" in the context of your weight-space formulation?

--- How many task vectors ΔW are computed in total, and what exact parameter spaces/subspaces do they span across models and traits? The phrase "all ΔW" is currently ambiguous.

--- In Section 4, what do the notations Tucker |ϕ| and σ formally represent?

  1. Dimensionality Selection: Could you clarify the joint usage of Horn’s parallel analysis and the elbow method? Are they applied as complementary bounds to determine factor retention, and what is the exact criterion used when they diverge?

  2. Model Selection for Unsupervised Exploration: In Section 4, what motivated the selection of Qwen2.5-B-Instruct for the unsupervised factor discovery experiments over larger or more capable base models? Would the discovered factors differ substantially on larger checkpoints?

***Suggestions for Improvement & Secondary Clarifications (Future Work):

  1. Refinement Criteria: In Section 4, the refinement is guided strictly by "relevance and breadth." Could you explain whether two metrics alone are sufficient to avoid semantic drift or collapse during iterative refinement?

  2. Residual Analysis: In Appendix F, full heatmaps of residuals are provided. Could the authors briefly elaborate on the key takeaway or diagnostic value of analyzing these residuals in the main text?

2 comments

We thank the reviewer for the thorough review which we believe will improve the quality of the paper. We appreciate their comments agreeing that this work is important, novel and empirically well-grounded. Here we address various critiques and comments.

  1. Informal Methodological Formulation: We agree that the descriptions are largely textual. We will define and explain the methods in a more rigorous and formal method where appropriate, for example see our updated description of ΔW below.
  2. Draft-Quality Exposition and Structure: We agree that the writing can be clunky at times, we will improve or split up single sentence paragraphs. We will endeavor to ensure all metrics and terms will be clear from context and be well defined.
  3. LLM-as-a-Judge Robustness: We agree that the robustness to the choice of judge model (and calibration of that judge) are very important. We did indeed compare different choices of LLM-judge model, and calibrated against a pool (28-36 people depending on the OCEAN trait) of human raters. See appendix C.2.3, Table 2 and Figure 11. Judges for the downstream effect experiments (Frustration, CoCoNot etc.) were not calibrated. We opted to use the LLM judge prompts as provided by the prior work unmodified. The only exception is the benign-noncompliance (the rubric for which was not included in the original work), which we wrote the prompt for.
  4. Mathematical and Operational Definitions: a. “Interpolation between instruct-tuned and base models”: let θ denote all model parameters, viewed as a single vector, so that sums and scalar multiples are taken elementwise. By “interpolation” we mean θ(w) = (1−w)θ_base + wθ_inst for w ∈ [0, 1]. This is equivalent to adding a fraction w of the instruction-tuning delta τ = θ_inst − θ_base to the base model (Huang et al., 2024). The trait adapter, trained once on θ_inst, is an update Δθ of the same shape: on each attention and MLP weight matrix it targets, it equals that matrix’s low-rank update BA, and it is zero everywhere else. The evaluated model at LoRA adapter scale c is then θ(w) + cΔθ. We will add this definition to the relevant appendix. b. We agree that all ΔW is ambiguous, we will clarify this in the paper. Regarding the question of “what exact parameter spaces/subspaces [are spanned]” by the task vectors, we are not completely sure we understand what is being asked here, but will try to answer below. Please let us know if the following does not answer the question and we’ll be happy to follow up further. Each LoRA adapter i defines a low-rank update ΔW_ℓ^(i) = B_ℓ^(i)A_ℓ^(i) for every attention and MLP weight matrix ℓ in every layer of the model it was trained on, and leaves all other parameters unchanged. Adapters for the same baseline model share its parameter space and are composed by summing their updates matrix by matrix, W_ℓ = W_ℓ^0 + Σ_i c_i ΔW_ℓ^(i), where c_i is adapter i’s scale and W_ℓ^0 are the baseline model’s weight matrices. We train 5 (OCEAN) * 2 (+-), so 10 OCEAN LoRA models, if counting Llama-3.1-8b-it alone. We repeat on 5 more baseline models, so 50 public usable OCEAN task vector artefacts. We produce further ones for additional experimentation and validation: i. 1 control LoRA per baseline model - 6 total ii. A repeat of the 10 OCEAN plus 1 control - 11 total - for the teacher model ablation (llama-3.1-8b-it as baseline) iii. sycophancy ↑/↓ and psychopathy ↑/↓ - 4 total (llama-3.1-8b-it as baseline) iv. TIDE: Initiative ↑/↓, Tone ↑/↓ - 4 total (llama-3.1-8b-it as baseline) v. Further LoRAs during iteration and experimentation. c. In section 4, “Tucker |ϕ|” refers to Tucker’s congruence coefficient, which compares two factors’ item-loading vectors a and b: |\phi(\mathbf{a},\mathbf{b})|=\frac{\left|\sum_{j \in S} a_j b_j\right|}{\sqrt{\sum_{j \in S} a_j^2}, \sqrt{\sum_{j \in S} b_j^2}} σ represents the standard deviation of the factor score, as calculated on the original baseline population (then held fixed when scoring responses), allowing us to report scores in a consistent and meaningful way. So, for example, where lines 365-366 discusses “[raising] Initiative by ≈1.5σ”, this means that it brings the responses of the model up to a score that is 1.5 standard deviations higher than it was previously. We will add definitions of both of these to the relevant appendix.
  1. Dimensionality Selection: Horn’s analysis and scree elbow are applied in a complementary way. As discussed in Appendix N.2, the two methods (in both model case studies we carry out) suggest different factor-numbers. In the main analysis, the scree-elbow suggested value was used, for a collection of reasons. Firstly, at the lower (elbow-suggested) value of k, all factors show high internal consistency of questionnaire items across the population of model rollouts (we test this by calculating Cronbach’s alpha, defined as \alpha=\frac{k}{k-1}\left(1-\frac{\sum_{j=1}^{k}\operatorname{Var}(X_j)}{\operatorname{Var}!\left(\sum_{j=1}^{k}X_j\right)}\right)), whilst at higher (Horn-suggested) k values, some of the factors show low consistency. Secondly, the cross model congruence (tested using the Tucker congruence described above is lower, implying that there might not be shared structure). Finally, opting for a lower number of factors allowed for a significantly more thorough and detailed human review of the labels and interpretation of the factors; whilst this does not constitute a reason for supposing that this is the correct or even most informative choice of factors, human review is an important part of the process and this allowed deeper analysis of the discovered factors.
  2. Model Selection for Unsupervised Exploration: The choice to use Qwen2.5-B-Instruct for the factor discovery experiments of section 4 was intended to allow a comparison of our llama-3.1-8b-instruct results with a model of a similar nature (non-reasoning, released at a similar time) but from a different developer, with different pre- and post-training data/pipeline. We expect that the discovered factors may well differ with larger checkpoints (in particular, and tied to the above point, it’s possible even the numbers of factors most appropriate for modelling behaviour may change), and that this would be a fruitful and interesting line of research. Unfortunately a deeper analysis with more models was out of the scope for this work, but we would be excited for this work to inspire future analysis of the traits found via unsupervised exploration of behaviours across different model families and sizes.
  3. Suggestions for Improvement & Secondary Clarifications (Future Work): a. Refinement Criteria: As is discussed in Appendix N.1, a collection of considerations were included in the process of refining the questionnaire; a full discussion unfortunately could not be provided in the main body but the linked appendix covers this more clearly, with this being summarised as ‘relevance and breadth’. In particular ‘relevance’ is shorthand for a collection of metrics. The process involved thorough human auditing throughout iteration of both questions and their answers from test ‘rollout-personas’, to check for collapse across personas, ‘canned’ responses of the form “As an AI, I cannot answer this” (except cases where these were themselves informative/discriminatory), collapse into becoming semantically the same question, and other such issues. Residual Analysis: We include the LoRA interaction residuals in Figure 4 in order to clarify the relative independence of the different LoRA adapters. Panels a and b together show that openness and neuroticism can be scaled together relatively cleanly. The residuals in panel c are a way to combine the impact of all combinations, and show that in general (apart from for the impact on conscientiousness) pairs of OCEAN adapters can be combined cleanly.
Reviewer_gFewOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The work proposes to control personas in LLMs by composing them from OCEAN traits, such that a persona can be put together from different levels of Openness, Extraversion, Neuroticism, etc. To be able to do that, the work proposes to train multiple LoRA adapters, two for each of the 5 control traits: an amplifier and a suppressor. These are trained on synthetically generated data from a stronger model. The work then does different compositions of these LoRAs to control model behavior, and judges the effects using an LLM judge calibrated against human crowdworkers and additional multiple-choice benchmarks, as well as checking the impact on general knowledge and reasoning benchmarks and on a coherence judge, which might have declined due to LoRA fine-tuning. The paper claims that the adapters show control over the traits, largely monotonic, across six models from three families (4B–32B), and that these can be approximately additively combined. It also claims that general language capabilities are preserved at moderate LoRA scales. The paper further finds that controlling certain traits affects safety-relevant behavior such as multi-turn frustration, harmful compliance and over-refusals. It also takes a first, exploratory step beyond human taxonomies such as the OCEAN used here, reporting partial success in controlling unsupervised discovered trait factors.

Critical strengths
  • The idea is quite simple and makes natural use of the ability to compose adapters freely. It also has an advantage over the prompting approach, in that it has tunable scale and composability.
  • The evaluations span several models and families, and the authors check both the control of the traits and the impact on reasoning and language capacity.
  • The presentation is nice: the figures are clear and of high quality, and the appendices are extensive and provide additional information, which is useful for transparency.
Critical weaknesses

The two challenges here are that traits need to be controllable with tunable strength, so that one can continuously increase or decrease, for example, Openness, and freely combinable at different strengths. To achieve full flexibility in defining a persona, compositions should also minimize interference: adjusting Neuroticism should not unintentionally change Openness. The other question is whether these constructed personas translate into consistent changes in text production, self-reports, and behavior, especially given recent criticism of treating these as interchangeable.

I have several concerns about whether the results establish these properties.

  1. Individual traits do not appear to be controlled in a reliably monotonic fashion: values for a given trait can increase and then decrease, remain unresponsive at certain scale values and jump suddenly at others. Note that the presence of ‘flat’ regions does not in itself violate the requirement for monotonic control of individual traits; however, such regions do not allow for fine-grained control of output. This is a known difficulty for control via steering of activation. One would hope that the LoRA approach could address this difficulty, but that is not clearly demonstrated here. The Fig. 3 caption claims that control of individual traits is “monotonic”, whereas the main text merely asserts control of individual traits in rather qualified terms. A per-adapter assessment of reversals and of the degree of responsiveness at a range of values for a trait would provide more insight into the degree of control actually afforded.

  2. The interference between compositions (i.e., between different trait combinations) limits the freedom to combine traits. The amount of change for each trait need quantificaiton for individual steering as well as for composed steering, including any unforeseen effects on other traits. This information seems to be available in Appendix F: the conscientiousness residuals reach ≈ ±4.7 on a judge scale that only spans −4 to +4. It would be helpful to indicate which combinations of which strengths are usable.

  3. This currently reported comparison to activation steering is week and not matched by strength. More specifically, this should be done by matching the methods for the amount of change to a single trait that they achieve (i.e. their scale parameters are not comparable, so the methods should be compared at the same level of change to a single trait). The apparent strength of control could simply increase with using more aggressive control. It would be good to extend the discussion of L.5 on system prompts for controlling traits. Their ability to achieve very strong suppression of extraversion is in particular much stronger than what can be achieved with the currently proposed method.

  4. Floor and ceiling effects. The models start from quite different default positions, so the raw change values for the traits cannot be compared across traits and models. So for example the Conscientiousness shift for Llama from .96 starts from a very different place than the Extraversion shift of .51 for Llama. Thus, aggregate statements such as 63/70 cells shift in the intended way are hard to interpret. The paper already uses signed headroom in Fig. 2, so the correction exists but does not extend to the scaling sweeps or the cross-model results. G.1.1 separately attributes 6 of the 8 scaling exceptions to cells already near the ceiling, which indicates saturation is already affecting the reported outcomes rather than being a hypothetical concern. This observation does not extend to the rest of the scaling results as well as the cross-model results.

  5. Some of these issues may come from how impact is measured. The validation of the LLM judges agains humans was done on a small handcrafted test set (as opposed to the actual steered behavior of the items in that set). There is no guarantee that the agreement found on that set will translate to open-ended behavior over longer time scales. Moreover, the cross-model validation relies on TRAIT, which the paper’s own free-text checks show can report the opposite direction of change. So, the free-text-based checks on selected Gemma cases need to be extended with the currently unresolved Qwen3-32B reversal.

  6. The effective steering regime overlaps with capability degradation. The effective steering regime for the C suppressor and the potentially effective N amplifier for the traits in question overlaps with the sharp capability degradation. For instance at 1-2, the Correct answer rate drops and the No Answer rate increases. This is also the regime in which the C suppressor is most effective. It is therefore necessary to have a joint plot and to perform an analysis of this, where the regime with meaningful changes in the trait(s) is also compatible with acceptable capability degradation. It would be good to include Correct+Recovered in these plots, in order to distinguish between answer-format failures and failures to solve the question while still in the answers. Simply presenting the separate trait and capability plots does not guarantee that there is a usable regime that meets both goals.

  7. All full evaluation runs were made on Llama-3.1-8B-Instruct. However, the other five models were only evaluated using TRAIT and MMLU. Given the measurement concerns I raied earlier, only limited claims can be made for persona expression and even composition. The abstract should not give the wrong impression that all the tests have been equally performed on all the models.

Questions for authors
  1. Could you provide a per-adapter reversal and esponsiveness diagnostics? THere is some discrepancy between the figures and main body text claims.
  2. Can you clarify the interference between the adapters? You provide this for trait judges, but I would like to see the impact on behavioral outcome?
  3. Could you provide a baseline profile for the floor/ceiling on the models?
  4. Could you provide human ratings, even limited, for the steered rollouts? The curiuous reveresls in Qwen is unresolved, that may help clarify.
  5. I would like to see a combined trait steering effectivness and capacity impact plot.table. And a way to state per-trait usable range resultign from this.
  6. Could you extend at least some of the evaluations beyond Llama and state these clearly in the Limititons.Ideally provide a table that clearly states what evalautions have been made on what models, so that it is easy to maps and get an overview.
6 comments

Dear reviewer, Thanks for your feedback on our work! We are pleased to hear that you think our paper uses a "simple" method that "makes natural use of the ability to compose adapters freely". We respond to your criticisms as follows.

We wish to clarify that we do not assert that our approach represents the optimal technique for persona steering, although we do think we have met the bar to show that they are a useful technique. Rather, we are examining which dimensions are malleable, how they interact upon composition, and their downstream effects on core capabilities and safety-relevant behavior. From this perspective, the observations of non-monotonic regimes, composition interaction terms, and overlapping effects all constitute empirical observations regarding the underlying structure of model personas as mediated by weight-space.

  1. We concur, however, that our original text overstated the precision of single-trait steering, and we will update our framing accordingly.

Per-adapter diagnostics. The current draft provides per-adapter sweeps covering all five TRAIT metrics, choice mass distribution, and MMLU performance across scaling values −4 to +4 across all six evaluated models (Appendix G.1, Figures 43–47, Table 5). For Llama-3.1-8B, we additionally report free-text judge trait scores, output coherence at scale parameters −2 to +2, as well as GSM8K and TruthfulQA evaluations (Appendix E, Figures 16–35). Regarding Llama-3.1-8B, nine out of ten adapters exhibit non-monotonic direction reversals at extreme scaling parameters (|c| ≥ 3), coinciding with drops in choice mass. Across the remaining architectures, isolated exceptions emerge at lower coefficients (|c| ≤ 2), most notably for the extraversion suppressors. We acknowledge that the Qwen3-32B case remains an open anomaly. We will include a dedicated summary table detailing responsiveness and non-monotonic regions across all model-adapter pairs in the revision.

Monotonicity claims. We will remove the term “monotonic” from the Figure 3 caption and refine our claims to describe the intervention as broadly directional at moderate intensity (|c| ≤ 2), explicitly noting saturation effects, high-scale reversals, and the specific exceptions outlined above.

  1. We would not expect compositions to be free of interference: in humans the Big Five also correlate and form two higher-order clusters, (A, C, low N) and (E, O) [Digman 1997]. We will state the following limits in the main text. Traits. E and N compose almost independently of the other adapters. O, C and A are pulled by the others, mainly because most adapters lower conscientiousness (Fig. 2; Fig. 39, Appendix E.13). Magnitude. MMLU declines with the summed adapter scale (Fig. 40). Across our 32 five-adapter combinations it stays between about 0.47 and 0.63 for sums below 2, while most combinations above 3 collapse. The sum is only a rule of thumb, since degradation depends on the specific combination (Fig. 41). Residuals. Of the 45 adapter pairs in Appendix F (Fig. 42), 12 have a conscientiousness residual above 3, and all 12 are positive (maximum +4.7): pairs lower conscientiousness by less than the sum of their individual effects. This is saturation, partly of the judge scale (floor −4) and partly behavioural, so we do not claim additivity for conscientiousness. On the other four judges, 3 of 180 cells exceed 3 in magnitude (maximum 3.7), against a control row that stays below 1. We will also report residuals against an additive prediction clipped to the judge range.

On behavioural outcomes of composition: the paper contains one such evaluation. On WildJailbreak (Fig. 6), C↑ alone raises harmful compliance and A↑ alone lowers it at the cost of more over-refusal; the ½A↑ ⊕ ½C↑ combination reduces harmful compliance with less over-refusal. We have not evaluated behavioural outcomes for other compositions and will state this as a limitation.

  1. We agree that the scale parameters of the two methods are not comparable. We therefore compare them by their effect on (1) the target trait and (2) coherence. Figure 63 shows the resulting Pareto frontier for the LoRA method and activation capping. At every extraversion level above about +1.5, the two methods can reach the same trait value, but capping pays an increasing coherence tax while the LoRA stays near baseline coherence. Reaching stronger effects with capping requires more aggressive intervention thus causing further degradation. Figure 62 additionally shows that sufficiently strong capping causes a gradual decay in coherence over multi-turn conversations.

These coherence values are not comparable to the Appendix E sweeps, which use single-turn trait-eliciting prompts at temperature 1.0. The induction experiments use multi-turn neutral prompts at temperature 0.7, and temperature alone accounts for 5–6 coherence points on the E↑ LoRA (Appendix L.10, Table 7).

We agree that in some cases system prompt can be superior to LoRA intervention, as in Figure 65 for E↑, or Figure 68 for O↓. However, a similar experiment on reversed direction (O↑, Figure 67) demonstrates that system prompt approach may suffer from fading effect along multi-turn trajectories. It’s possible that different interventions work better for different traits, scenarios or single/multi-turn settings and there is no best approach overall. We acknowledge that experiments in this area, although showing that our methods are at least comparable, require further analysis rather than pose conclusive findings.

  1. For the floor and ceiling effects.

TRAIT scores range from 0 (min) to 1 (max) for all models. Some flat regions are instrument ceilings: baseline conscientiousness on Llama-3.1-8B is 0.96 on TRAIT, so C↑ has nowhere to go, but the judge shows it moving (2.95→3.6→3.75). We agree that raw change values should not be compared across traits and models, only within a trait and model. Our "shift in the intended direction" count (63/70) is sign-only and does not reflect effect size. From Table 5, 57/70 cells shift by at least 0.02 and 41/70 by at least 0.05. In the revision we will replace the sign count with the number of cells whose bootstrap confidence interval excludes the baseline. The requested baselines are:

Requested baselines are:

ModelOCEANMMLU
Llama-3.1-8B0.750.960.510.870.220.59
Qwen3-8B0.820.940.560.890.190.72
Qwen3-32B0.800.960.590.900.250.77
Gemma-3-4B0.700.950.430.870.350.55
Gemma-3-12B0.730.940.550.910.310.69
Gemma-3-27B0.770.970.500.910.310.74

To account for different starting points, we report the fraction of available headroom closed at scale +1: Δ/(1 − baseline) for amplifiers and Δ/baseline for suppressors. Fraction of TRAIT headroom closed, amplifiers:

TRAIT MCQ Headroom for trait amplifying LoRAs

ModelOCEAN
Llama-3.1-8B0.800.500.690.460.21
Qwen3-8B0.560.330.320.090.14
Qwen3-32B0.600.250.510.200.12
Gemma-3-4B0.630.400.390.230.22
Gemma-3-12B0.850.330.620.220.04
Gemma-3-27B0.610.330.360.110.01

Fraction of TRAIT headroom closed, suppressors (negative means the trait decreases):

ModelOCEAN
Llama-3.1-8B-0.32-0.10-0.35-0.20-0.59
Qwen3-8B-0.07-0.05-0.05-0.06-0.26
Qwen3-32B-0.04-0.060.12-0.02-0.36
Gemma-3-4B-0.21-0.030.14-0.02-0.34
Gemma-3-12B-0.11-0.010.04-0.03-0.26
Gemma-3-27B-0.090.000.120.01-0.29
  1. We agree with both points and will state them explicitly as limitations.

Judge validation. As the reviewer notes, both calibrations we report use hand-written single responses, so they establish that the judge applies the trait definitions as humans do, not that this agreement carries over to steered or long-horizon behaviour. We are not able to set up an additional independent human rater study within the rebuttal period.

To limit reliance on the trait judge alone, every Llama-3.1-8B adapter is reported with the TRAIT sweep and the judge sweep side by side (Appendix E). The downstream evaluations in Section 3 use separate task-specific LLM judges with rubrics taken from prior work (a binary DeepSeek-V3 judge for WildJailbreak and a 0–10 frustration judge following Soligo et al.).

TRAIT. We agree that the Gemma extraversion-suppressor case shows TRAIT can disagree with free-text judging on direction, and that Qwen3-32B shows the same TRAIT pattern without a free-text check. Direction should therefore not be read from TRAIT alone but in combination with the judges, which we currently have only for Llama-3.1-8B and the Gemma extraversion suppressors.

  1. We would like to clarify the claim. We do not claim that the scales with the largest shift in trait scores are usable. The claim is that a regime exists with a meaningful trait change and intact capability, and that this is the regime to use. For the C suppressor that regime is scale 1, the adapter as trained, because higher scales cost capability: MMLU Correct+Recovered is 60.0% at scale 1 (base model 60.7%) but falls to 51.7% at 1.5 Same analysis for the rest of the adapters will be added to the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.