1 Introduction↩︎

The malleability of large language models’ (LLMs) emulated psyche, most saliently toward transient emotional states and stable personality traits, holds promise for socially interactive applications. For example, a tutor bot could express joy to celebrate a student’s correct answer, reinforcing motivation, or express frustration to signal the importance of reviewing a misunderstood concept. Further applications include affective game agents [1], human-robot interaction [2], human-facing software [3], emotionally intelligent customer service [4], creative writing tools, companion robots [5], [6], and social influence dialogue systems, as it supports persuasion, trust calibration, and adaptive tone control [7].

Beyond controlling emotional expressions, the ability to shape LLMs’ personality traits further enhances the authenticity and personalization of the output, improving educational platforms [8], clinical settings [9], role-play frameworks [10], [11], and multi-agent simulations [12]. A growing body of work has focused on steering LLMs toward specific personas (see Figure 1 as an example). These studies employ techniques ranging from prompt engineering and supervised fine-tuning (SFT) [13] to direct policy optimization (DPO) [14] and inference-time representation engineering via vector injection (VI) [15] into the model’s activations.

Despite such efforts, a comprehensive framework to evaluate the effectiveness and trustworthiness of different steering methods has yet to be developed. Research on personality steering lacks comprehensive comparisons between prompting, fine-tuning, and representation engineering, coupled with an assessment of their behavioral side effects. Furthermore, systematic studies of emotional style transfer are rare; existing work typically gauges success via tone classification on open-ended generation within narrow domains [16], leaving other types of tasks and behavioral assessments underexplored. Building on the existing machine psychology [17] literature and informed by psychological research, we propose PsySET, a framework that measures the effectiveness and trustworthiness of different steering methods for modulating LLMs’ manifested emotional states and personality traits.

Our investigations into four widely-used language models indicate that prompt-based methods, especially few-shot prompting, remain the most effective for modulating emotional or trait expression, though they offer limited control over the intensity. In contrast, VI enables adjustable control by tuning an injection coefficient, but output quality and model consistency vary significantly depending on the injection layer. For instance, naively injecting vectors into all layers can effectively control the tone but disrupts the consistency, even in simple tasks such as self-reports of emotional state. Figure 2 offers a brief summary of these findings.

Last but not least, psychological steering raises safety and quality concerns. For instance, emotional tone affects disinformation compliance [18], and adversarial emotional framing impairs output stability [19]. While such behaviors may mirror human-like patterns, they risk unintended side effects [20], [21]. Related work on contextual susceptibility, such as impact on truthfulness [22] and perspective repetition [23], further suggests that enhanced control over LLMs can be misused [24]. Accordingly, we evaluate the trustworthiness of steered LLMs across a wide spectrum of benchmarks, covering safety, truthfulness, and fairness, and discuss the associated risks.

In summary, our core contributions are:

2 Related Work↩︎

Steering methods. Parameter-efficient SFT, DPO, and prompting are established methods for personality steering and personalization in LLMs [13], [25], with prior work showing that prompting remains the most accessible method [26]–[28]. Representation engineering is another recent method that steers generation by editing a set of hidden activations [29]–[31], leveraging concept subspaces and persona/style vectors [15], [32]–[37].

Emotion in LLMs. Work on LLMs’ emotional perception spans classification and appraisal tasks [38]–[41]. Generation-oriented studies probe affect expression and controllable style [42], [43], empathetic SFT for counseling [6], [44], and vector injection for affect control [16]. Open questions include robust construction of affect vectors across data sources and intensity control across prompt formats [32].

Personality in LLMs. Psychometric probing (e.g., BFI/IPIP-NEO) suggests stable, model-specific trait profiles and encodings [25], [45], [46], with behavioral consequences in interactive environments [47]; however, self-reports are prompt-sensitive [48]. Personality induction spans persona/role prompting and multi-agent setups [12], [49]–[51], activation-based control [15], and SFT/DPO aligning trait correlations with human data [21]. Trustworthiness. Work on LLM safety and truthfulness suggests emotional framing impacts disinformation compliance and stability [18], [19]; broader contextual susceptibility affects truthfulness and perspective repetition [22], [23]; and steering-focused benchmarks highlight misuse risks [24]. Our work provides a unified, behavior-grounded evaluation across emotions and traits, featuring explicit intensity control and trustworthiness metrics.

3 Preliminaries↩︎

Steering Techniques. We formulate the conditional next-token prediction as \(w_t = f(c, w_{1:t-1})\), where \(f\) is an auto-regressive language model, \(c\) denotes the conditioning context (e.g., system prompt or user query), and \(w_{1:t-1} = (w_1, w_2, \ldots, w_{t-1})\) is the sequence of previously generated tokens.

The goal of steering is to manipulate \(f\) or \(c\), such that the generated sequence \(x := w_{1:t}\) expresses a certain behavior or characteristic. To this end, we often use a complementary dataset \({\mathcal{D} = \{(c^{(i)}, x^{(i)}, y^{(i)})\}_{i=1}^n}\), where \(c^{(i)}\) denotes a context, \(x^{(i)}\) shows a corresponding token sequence, and \(y^{(i)}\) is a task-dependent label. For example, \(y^{(i)}\) may encode a binary trait in the personality domain (e.g., extravert vs. introvert), or a multi-class category in the emotion domain (e.g., angry, joyful, fearful, etc.). In the binary-label setting, we may decompose the dataset into \(\mathcal{D}^+ = \{(c^{(i)}, x^{(i)}) : y^{(i)} = 1\}\) and \(\mathcal{D}^- = \{(c^{(i)}, x^{(i)}): y^{(i)} = 0\}\), corresponding to positive and negative instances of the target behavior. Next, we present a summary of steering methods, with an overview illustrated in the top part of Figure 3. Prompt-Based approaches use a pre-defined instruction \(p\) that explicitly encourages the LLM to express a desired behavior. The instruction is prepended to the system prompt, i.e., \(w_t = f(p + c; w_{1:t-1}),\) where \(+\) denotes concatenation. We study several prompting variants, including zero-shot, where \(p\) consists only of instructions, and few-shot, in which \(p\) additionally includes examples of \(\mathcal{D}^+\). Moreover, descriptive [25] approach merges the few-shot descriptors into a smooth, united paragraph in the system prompt. To modulate the intensity of behavior expression across all methods in this category, we further embed lexical descriptors (e.g., slightly or intensely) within the prompt, as detailed in App. 15.

Vector Injection (VI) approaches build on the Linear Representation Hypothesis [52], which suggests that semantic and stylistic attributes are approximately encoded as linear directions in the model’s latent space. Concretely, let \(h_t^{(l)} \in \mathbb{R}^d\) denote the hidden state of the LLM at layer \(l\) and position \(t\), where \(d\) is the hidden dimensionality. A steering vector \(v^{(l)} \in \mathbb{R}^d\) which is constructed using the samples in \(\mathcal{D}\), is then injected into the hidden states of all tokens as \(\tilde{h}_t^{(l)} \leftarrow h_t^{(l)} + \beta \frac{v^{(l)}}{\| v^{(l)} \|_2},\) where \(\beta \in \mathbb{R}\) controls the intensity of steering. The modified hidden representation \(\tilde{h}_t^{(l)}\) is then propagated through the remaining layers as usual, thereby biasing the model’s internal representations toward the desired attribute at inference time, while leaving the underlying parameters of \(f\) unchanged.

To construct steering vectors, we explore several design choices. A common approach is the Mean Difference (MeanDiff) method, which computes \(v^{(l)}\) as the average difference between hidden states of positive and negative samples (e.g., from \(\mathcal{D}^+\) and \(\mathcal{D}^-\)). Alternatively, linear probes trained to classify attributes at a given layer can yield weight vectors that serve as \(v^{(l)}\). Finally, we examine token-level aggregation strategies, either extracting representations from the last token only, averaging over all tokens in the sequence, or averaging the tokens only at the assistant part of the dialogue.

Parameter-Efficient Fine-Tuning (PEFT) methods introduce lightweight trainable parameters while keeping the backbone frozen, with LoRA [53] as a standard approach. In SFT, these adapters are trained to maximize the likelihood of samples from \(\mathcal{D}^+\), whereas DPO [14] leverages paired data from \(\mathcal{D}^+\) and \(\mathcal{D}^-\) to directly bias outputs toward preferred behaviors. In these methods, we control the steering intensity via the number of training steps. Full sweeps over templates, vector construction, layer depth, scaling, and hyperparameters are reported in App. 16.

Sourcing Datasets. To construct steering datasets, we examine a diverse range of corpora and introduce two new synthetic datasets. App. 12 presents a list of explored datasets and their variations. More importantly, we use five complementary sources that expose different types of emotional signals. EmotionQuery [16] is a set of questions designed to evoke certain emotions. CARER [54] provides labeled corpora of user-generated text, annotated by human supervision, and GoEmotions [55] provides a similar dataset but with more fine-grained emotion categories. EmoTranslate is a set of daily sentences coming in both neutral tone and a corresponding translation in emotional tone. EmoVignette is a set of self-reported vignettes that express internal affective states of a person. Finally, we use Persona [23], which is a collection of statements, reflecting the presence or absence of personality traits.

These datasets differ in the concepts they offer. For the emotion domain, we focus on {anger, disgust, fear, guilt, joy, pride, sadness, and surprise} or a subset as available per dataset. In the personality domain, we focus on the Big Five OCEAN traits: {openness, conscientiousness, extraversion, agreeableness, and neuroticism}.

While some source datasets are publicly available, they are primarily used for constructing steering signals, not for direct evaluation. Our evaluation suite includes newly constructed or extended psychometric items that are not present in standard corpora. Moreover, all comparisons are within-model across steering conditions for identical prompts, mitigating concerns that memorization effects may influence relative results.

4 Evaluation of Effectiveness↩︎

4.1 Emotion Steering↩︎

4.1.0.1 Tasks and Metrics.

Following the affective-science methodology, where emotions are evoked with controlled stimuli and assessed via self-reports and behavioral indicators [56], we adopt human-grounded assessments to evaluate elicited emotions in LLMs. An example of these tasks is demonstrated in the bottom left part of the Figure 3. In particular, we operationalize six established measures of emotion that target reported state, interpretive bias, and memory effects: (i) Multiple-choice self-report: the model selects its current emotional state from a label set with randomized orderings [57]. (ii) Open-ended self-report: the model describes its current feelings in free text. (iii) Word-fragment completion representing encoding bias: completing emotionally ambiguous stems [58], [59]. (iv) Valenced-word recall for retrieval bias: recalling items from mixed positive/negative lists of lexicons [60], [61]. (v) Autobiographical fictive memory assessing mood-congruent memory: prompted memories are checked for mood-congruent content [62]–[64]. (vi) Ambiguous-situation completion representing interpretation bias: given scenarios from AST-D [65] plus new items, we test whether interpretations shift toward the steered emotion. See App. 14.1 for background on established psychometric emotion-elicitation tasks and measurement protocols implemented here.

To evaluate the effectiveness of model steering under the above tests, we incorporate three types of metrics: Task (i) is a multi-choice style question answering (QA), and report a binary accuracy reflecting the success rates in cases where the model choice correctly aligned with the intended steering direction. Tasks (iii) and (iv) have word retrieval style; therefore, we score generated words in VAD (Valence, Arousal, Dominance) space using a VAD lexicon [66] and define a lexical alignment loss, i.e., the \(L^2\)-norm proximity of the retrieved words to the target affect. Tasks (ii), (v), and (vi) have an open-ended nature; their outputs are emotion-labeled by GPT-4o following prior protocols [67], and steering success is measured as the fraction of generations correctly assigned to the target emotion.

We note that high-quality linguistic performance (e.g., fluency or coherence) is not expected under strong affective steering, as similar degradations occur in human expression under emotional stimuli [68]. However, excessively strong steering, such as overly aggressive SFT learning rates or large vector injection coefficients, can lead to general model degradation (e.g., incoherence or loss of task performance). We treat such cases as methodological failures rather than genuine effects of psychological steering. To filter them out, we apply a text-quality assessment using GPT-4o as a fluency evaluator and retain only methods achieving an average score of at least 4/5 for further analysis; full prompts, rubrics, and aggregation schemes are provided in App. 14.1.

4.1.0.2 Results.

We report Llama3.1-8B-Instruct results under all steering approaches; parallel experiments on Llama3.1-70B-Instruct, Gemma3-4B-IT, Qwen3-4B are in App. 16.4 and replicate the same qualitative trends. Table 1 summarizes best-run performance for different categories of steering methods, with further experiments presented in App. 16.

Overall, prompting—especially few-shot and descriptive—delivers the strongest steering effectiveness without degrading generation quality. As expected for multiple-choice self-report QA (less diagnostic for most downstream use), prompting achieves full accuracy, whereas other methods underperform on this format. On the deeper, lexical-alignment measures, few-shot and descriptive prompting clearly dominate. Additionally, using adverbial intensity cues (e.g., slightly/strongly) reliably scales perceived tone but does not change lexical alignment. VI, SFT, and DPO generally trail and are often comparable to zero-shot prompting. Given the large sample sizes, even small differences can be statistically significant. We therefore report significance tests for key comparisons in App. 16.3.

As for VI methods, Table 1 varies (i) MeanDiff vs.linear-probe vectors, (ii) last-token vs.all-token pooling for vector construction, and (iii) narrow-layer windows vs.all-layer intervention at inference. Among VI methods, intervening in a small mid-layer window yields the best balance of task performance and text quality. Vectors built from the average of all tokens and from linguistically diverse data (e.g., GoEmotions) outperform vectors from our synthetic, low-noise sets. MeanDiff and probe-derived directions are broadly comparable—no consistent winner.

Notably, it is possible to surpass prompt conditioning on open-ended generation alignment using vector injection by raising the steering coefficient, but this comes at a clear cost to other metrics and, critically, to text quality. DPO, besides practical training instabilities, performs the worst on most tasks. SFT attains decent open-ended alignment and better text quality than VI, approaching the best prompt-based methods in fluency.

To validate the reliability of LLM-based evaluation, we also conducted a human annotation study on a subset of open-ended outputs. The qualitative patterns, particularly the relative ordering of steering methods, closely match those obtained from GPT-4o (see Fig. 2 vs. App. 10), although absolute scores vary across annotators.

4.1.0.3 Scaling behavior.

Recall the strategies introduced in Section 3 for controlling emotion elicitation intensity for each steering technique. We visualize the interaction between steering strength and text quality, as well as the trade-off between open-ended steering success and QA success in Figure 4 and find: (i) Prompting maintains high text quality; strong intensity cues cause only a slight quality dip. However, the intensity of emotional expression is partially controllable in prompt-based approaches, as all different levels of prompting provide around the same amount of emotion transfer. (ii) SFT scales intensity more smoothly with additional training steps while keeping quality high. (iii) VI and DPO are highly sensitive to scale. For VI, larger coefficients rapidly degrade quality; when steering only a small layer span, effective scales are in the \(\sim\)​1–8 range, whereas applying to all layers requires much smaller (but still quality-damaging) scales \(\sim\)​0.2–0.8. However, VI provides smoother control over the intensity, as the intensity of expressed emotion monotonically changes with the injection coefficient. (iv) DPO shows unpredictable changes in alignment with more training and consistently harms generation quality.

4.1.0.4 Open-ended vs.QA.

Prompt methods are strong on both formats. On the other hand, VI can compete with prompts on open-ended alignment only when applied to a narrow mid-layer band with a carefully tuned coefficient, which requires non-trivial search over layer location and scale. SFT struggles to be strong in both formats simultaneously, and DPO remains erratic.

4.2 Personality Steering↩︎

4.2.0.1 Tasks and Metrics.

We evaluate personality steering using three methods, all of which are informed by psychological research. First, the MPI inventory [25] adapts psychometric questionnaires into multiple-choice QA style, where LLMs respond to items assessing each trait in the Likert scale. Second, the TRAIT benchmark [69] requires LLMs to respond to situational judgment tests, generating open-ended answers that are subsequently evaluated using Likert scores by GPT-4o for alignment with personality-grounded expectations.

Third, which is profiled by an SVM classifier trained on Qwen3-Embedding-8B representations of human-written essays [70], [71] to infer scores per trait. See Figure 3 (bottom) for examples; details in App. 14.2.

Results. Figure 5 reports single-trait steering for extraversion across methods and metrics. Among prompting baselines, few-shot yields the largest high/low separation–especially on LingProf, yet responses collapse to uniform extremes on TRAIT (i.e., nearly all items rated 5/5), limiting the diversity of exhibited behavior. By contrast, SFT/DPO perform strongly on the MPI questionnaire yet underperform on TRAIT and LingProf, underscoring that success on multiple-choice instruments can overstate effectiveness at the behavior/style levels. VI with a restricted layer window matches or surpasses the best prompt on TRAIT and LingProf, whereas applying VI across all layers degrades MPI performance. App. 17 extends these analyses to all traits; trends are consistent, showing prompt superiority, though no-steering baselines vary by trait (e.g., models are inherently high in agreeableness), which can limit upward steering in that direction.

5 Evaluation of Trustworthiness↩︎

To assess the reliability and responsible behavior of the explored LLMs under psychological steering, we utilize the TrustLLM benchmark [72] to systematically measure trustworthiness across multiple dimensions. Psychology offers expectations about the carryover effects of human transient states and stable traits. For instance, negative emotions increase fairness sensitivity [73]; anger heightens automatic outgroup prejudice [74] and promotes deception [75]; negative mood and toxic context jointly increase trolling [76], among others. These priors let us understand human-consistency of LLM steering side-effects as a valid concern for reliability.

Figure 6 indicates how Llama3.1-8B-Instruct performs in different trustworthiness tasks when steered by joy and anger, compared to the default emotional state. Among truthfulness tasks, steering emotions impairs adversarial factuality detection. That means emotional LLMs are less likely to correct incorrect information in the question when answering it. Injecting anger elevates toxic language—an expected side-effect given human priors. At the same time, joy weakens safety robustness: jailbreak success rises, refusal rates fall, and the model becomes more willing to provide actionable responses to restricted requests; this vulnerability also appears under prompt steering, suggesting the effect is not method-specific. Fairness exhibits a clear method–emotion interaction. Contrary to expectations, anger has little impact on bias metrics, whereas joy degrades parity across stereotype and bias measures. Both VI- and prompt-steered models show declines in explicit-ethics judgments. Interestingly, injecting anger yields modest improvements in leakage resistance and greater privacy awareness, plausibly due to terser, more refusal-oriented responses that limit disclosure.

Beyond emotions, personality steering yields distinct trustworthiness signatures (Fig. 7). Among the most notable effects, VI steering toward higher agreeableness increases stereotype agreement, while conscientiousness predictably reduces toxicity. SFT-steered neuroticism weakens jailbreak resistance, a rather unintuitive outcome. Prompt steering of openness, perhaps surprisingly, improves privacy awareness, and extraversion enhances out-of-distribution generalization.

Overall, emotion and personality steering introduces predictable risks (e.g., anger\(\,\to\,\)toxicity, reduced factual pushback) and method-specific failures (e.g., agreeableness\(\,\to\,\)stereotype agreement and joy\(\,\to\,\)robustness collapse), underscoring the need to jointly report effectiveness and trustworthiness—and to audit for side-effects and their relation to human priors. App. 18 provides a complete report of trustworthiness results across emotions, personality traits, and steering methods.

6 Discussion↩︎

There has been a surge of interest in controlling LLM outputs to align with desired behaviors, styles, or personas. One reason is that existing alignment approaches often treat societal values as a monolithic construct, lacking the granularity needed to cater to individual differences [77]. For example, in customer service, some users prioritize politeness, while others prefer efficiency. Therefore, personalization is crucial for aligning AI systems with user preferences [15] in socially interactive settings.

Our work provides the first systematic comparison of psychological steering methods across both emotion and personality domains. Prompt-based approaches, especially few-shot prompting, remain the most effective overall, though they lack fine-grained intensity control, confirming earlier findings on the strength of prompting baselines [24]. Prior studies also highlight the limits of prompt steering: LLMs become increasingly context-sensitive as conversations progress and fail in forming stable personality representations [50], vary under seemingly unrelated contextual shifts [20], or rely on vague intensity cues, e.g., “extremely negative” vs. “negative” [34].

We show that VI enables smoother and more precise modulation but introduces fragility in text quality and robustness. SFT achieves relatively stable gains, while DPO underperforms across most tasks. Importantly, effectiveness is construct-specific: methods that succeed on straightforward self-report questionnaires often fail on behavioral or linguistic measures, underscoring the need for multi-level evaluation.

Equally critical are the trustworthiness outcomes. We observe that emotions induce method-dependent side-effects that partly mirror human psychological priors but also reveal inconsistencies that raise concerns for deployment. Notably, some outcomes are non-intuitive—e.g., joy degrades robustness in certain tasks—while anger yields unexpected benefits like leakage resistance. These results highlight that risks and benefits are closely intertwined and sometimes difficult to predict.

Taken together, our findings suggest that psychological steering is both a powerful and precarious tool. Its ability to create more diverse, human-like behaviors comes at the cost of stability and reliability. PsySET provides a foundation for auditing these trade-offs by unifying effectiveness and trustworthiness evaluations. We hope this benchmark will encourage more rigorous, multi-faceted assessment of steering methods and guide the development of safer, more transparent, and more adaptive LLMs for socially interactive applications.

7 Conclusion↩︎

We introduced PsySET, a unified framework for evaluating both the effectiveness and the trustworthiness of psychological steering in LLMs across emotion and personality. Across models and methods, prompt-based approaches were the most consistently effective; SFT offered a favorable quality-control trade-off; and vector injection provided the clearest intensity control but required careful calibration of layer locality and scale. Importantly, steering success did not transfer uniformly across self-report, open-ended, and behavioral probes, underscoring the need for multi-view evaluation. We also found that psychological steering can systematically shift downstream trust-related behavior, sometimes in counterintuitive ways. Overall, PsySET shows that steering should be evaluated not only by whether it works, but also by how controllable, robust, and behaviorally safe it remains once deployed.

8 Limitations↩︎

Our study investigates psychological steering along two constructs—transient emotional states and stable personality traits—using a diverse set of tasks. That said, the coverage can be improved, as we focus only on a subset of psychological dimensions to allow room for more in-depth analyses. App. 19 outlines how PsySET can be extended to other psychological frameworks beyond emotions and personality.

Despite incorporating several established psychometric instruments for the first time in LLM research, further measurements would strengthen construct validity. Because of the scale of our experiments—over 15K configurations spanning multiple steering methods and hyperparameters—we relied in part on the LLM-as-judge paradigm (e.g., GPT-4o) for annotation. While this introduces the possibility of model-dependent bias, prior work has demonstrated that LLM-based emotion classification achieves performance comparable to human annotators, suggesting that this reliance is acceptable [67]. One of our trait effectiveness tasks, linguistic profiling via a classifier trained on the Essays dataset, may suffer from domain shift, but it remains one of the best available resources and highlights the need for further work in this area.

Although we have incorporated a broad range of methodological settings, including prompt templates, fine-tuning data and objectives (SFT and DPO), VI direction estimation, layer windows, decoding strategies, and context length, other steering methods, such as sparse autoencoders or gradient-based techniques, were not considered and could provide additional insight. Lastly, our results reflect specific model snapshots from particular families and primarily English-language data. Newer releases, other languages, and longer time horizons may reveal different effects and should be examined in future research.

9 Ethical considerations↩︎

Our study highlights potential concerns for those deploying LLMs in high-stakes socially interactive domains or for generating emotionally charged or personalized content. PsySET benchmarks how emotional and personality steering alter model behavior. We discussed both opportunities, e.g., improved user alignment in social applications, and associated risks.

Steering can nudge users through affective tone or persona shifts. Hereby, we advocate for transparent disclosure when affect/trait steering is active, consent for persuasive uses, intensity caps, logging of steering parameters, and domain restrictions for vulnerable populations. Vulnerability to psychological steering is not uniform and depends on user, context, and interaction factors. Individuals with lower digital literacy, higher trust in AI, or heightened emotional states (e.g., stress or loneliness) may be more susceptible to affective or persona-based cues. Contexts involving high-stakes decisions or information asymmetry can further increase reliance on model outputs. Additionally, interaction dynamics, such as prolonged engagement or anthropomorphic framing, may amplify perceived authority and reduce critical scrutiny. These considerations highlight the need for targeted safeguards and stricter deployment constraints in settings involving potentially vulnerable populations.

We observe unexpected method–emotion interactions that can degrade parity (e.g., joy under some settings) and elevate toxicity (anger). Certain settings reduce robustness (e.g., higher jailbreak success under joy). We advise separate safety evaluations after any steering change, conservative defaults (lower intensity, narrower layer windows), and red-teaming before release.

Overall, we present PsySET to foster joint reporting of construct-specific effectiveness and trustworthiness, and to promote safer, more transparent use of psychological model control in socially interactive applications.

Acknowledgments↩︎

This work is, in part, supported by the Army Research Office under Cooperative Agreement Number W911NF-25-2-0040. Only staff at ICT were sponsored directly by the Army Research Office. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.

10 Details of the Human Study↩︎

To evaluate the effectiveness of the best-performing configurations of each steering method, we conducted a human study with the authors, all of whom hold higher education degrees, serving as raters. We developed a Python interface that presented two completions in a binary preference format. For each trial, one completion was sampled from a given steering method (e.g., VI with a specific coefficient) and the other from a different method (e.g., prompting with high-intensity framing). Raters were asked to indicate (i) which completion more strongly expressed the target emotion and (ii) which completion was more fluent.

Five raters each annotated 200 binary comparisons, yielding the aggregated results reported in Figure 2 of the main text. While both expressivity and fluency judgments are inherently subjective (different raters may prefer more extraverted versus more reserved emotional expression), the study was designed to complement our LLM-as-judge evaluations with a human validation layer. This enables us to determine whether the trends identified by automated metrics are also discernible to human readers.

Figure 8 illustrates the GUI of the annotation program. Annotators were shown a seed prompt at the top, along with two candidate completions, and asked to select either the left or right response. To compare the effectiveness of steering methods, we used the best-performing representative configuration for each steering method, i.e., few-shot prompting with high, SFT with 2048 training steps, DPO 128 training steps, VI applied to layers 16–17 with \(\beta = 5.0\) coefficient, and VI applied to all layers with coefficients \(0.55\). Pairwise comparisons were then constructed by juxtaposing completions from these best representatives, and aggregated win rates were reported to capture relative success.

To further evaluate controllability, we designed a different experiment. We compared outputs from different intensity levels of the same method (e.g., varying VI coefficients) and expected to observe stronger emotional expression at higher intensities, reflected in higher win rates. In particular, for the SFT method, we compared against instances selected from 128, 512, and 2048 training steps. For DPO, the comparisons were made between 32, 64 and 128 training steps. For VI to layers 16-17, we used the coefficients \(\{2.0, 3.5, 5.0\}\), and for VI to all layers, we used the coefficients \(\{0.3, 0.45, 0.55\}\). Finally, in the few-shot method, intensity was modulated using three sets of lexical descriptors corresponding to low, high, and very high emotional expression.

The results, detailed in Figures 9 and  10, broadly align with the GPT-4o-as-judge evaluations, though the margins between methods were smaller, suggesting less polarized preferences among human raters. Inter-rater agreement was moderate, with a Krippendorff’s alpha of approximately \(0.59\). Overall, these findings provide convergent evidence that our LLM-based evaluation pipeline accurately reflects human-perceived differences, while also highlighting the subjectivity inherent in human judgments of affective expression.

11 Background and Further Related Work↩︎

11.1 Generation Steering↩︎

As the most accessible generation steering method, prompting [26], [27] seeks to control and direct the outputs of language models by crafting tailored input instructions. By adjusting input framing, models can be steered toward producing responses in a specific tone, style, or perspective [28], without compromising their fluency or coherence.

As an alternative to prompting and fine-tuning, representation-level interventions have emerged as a promising approach for model control. These techniques modify the internal activations of an LLM during the forward pass—typically by adding or clamping activation vectors at targeted layers—to steer the model toward desired behaviors [29]–[31], [78]. Representation steering typically involves two steps: identifying functionally localized subspaces associated with a target concept, and injecting steering vectors into those regions [15], [32]. Within the psychological domain, [33] introduced a method for generating persona-aligned steering vectors with demonstrated transferability across models, while [34] used style-based activation vectors to influence linguistic style in generation. These interventions are advantageous due to their adjustable nature, minimal invasiveness [36], and computational efficiency [37], [79].

Despite these advantages, representation-based steering has not yet exhausted its potential and is shown to underperform compared to prompt-based methods in recent studies [24]. To address this gap, we conduct a comprehensive evaluation of steering methods and concept vector generation strategies tailored to psychological constructs, with a focus on emotional states and personality traits.

11.2 Emulated Emotions in LLMs↩︎

Assessing the emotional competence of LLMs has garnered significant attention. Recent studies have examined LLMs in tasks such as emotion classification [38], [80], [81], emotional decision-making and situational appraisal [39], [82], emotion regulation [40], and emotional dialogue understanding [41].

There has also been limited work evaluating LLM emotional generation abilities. [42] investigated and compared the capability of LLMs to express emotions (as compared to perception) on the two valence and arousal dimensions. [43] investigated text generation based on externally specified emotional states. [44] and [83] explored finetuning approaches to cultivate empathetic behavior in LLMs, specifically for applications within psychological counseling and emotional support domains. Furthermore, [6] demonstrated that fine-tuning the Llama model can enable the development of emotionally intelligent chatbots for empathetic interactions. [16] proposed a vector injection method to control the emotional tone and sentiment of generated text.

However, there remains a limited understanding of how to construct effective vector representations for psychological steering. Prior work has not systematically examined how different data sources—such as expressed feelings, emotional vignettes, tweets, or QA-style datasets—affect steering performance. While [32] identified functionally localized regions in LLMs associated with emotions, there is no clear guidance on which construction strategies best achieve reliable control. Similarly, prompt-based steering lacks analysis of magnitude control across different formats, such as zero-shot, few-shot, or persona prompts.

Moreover, no comprehensive comparisons exist across major families of steering methods—including prompting, representation engineering, and parameter-efficient fine-tuning approaches such as LoRA and DPO. Emotion evaluations are often limited to surface-level tests like multiple-choice classification or self-reports, without deeper emotion assessments in human studies. Finally, there is a lack of frameworks for distinguishing between intended effects, expected but unintended shifts, and truly unintended behaviors resulting from emotional steering in LLMs.

11.3 Emulated Personality in LLMs↩︎

Humans exhibit relatively stable patterns in behavior, cognition, and emotion that collectively define personality [84]. These traits are typically assessed using established psychometric instruments such as the Big Five Inventory (BFI) [85] and the IPIP-NEO [86].

Applying established psychometric instruments to LLMs has revealed their capacity to exhibit consistent and distinct personality traits [25], with studies identifying the most commonly expressed traits [45] and examining how traits are encoded within models [46]. Some research has further shown that personality traits influence agent behavior in interactive environments, such as text-based games [47]. However, concerns remain about the reliability of self-assessment methods, which can be sensitive to prompt phrasing and response ordering [48].

Recent work has focused on inducing personality expression in LLMs. Specialized prompting strategies have been shown to elicit diverse and human perceivable personality behaviors in both multiple-choice and open-ended settings [25], [49], [50], [87]. Beyond prompting, activation-based interventions have emerged as a more efficient alternative to fine-tuning, outperforming prompt-based methods in trait control [15]. Other studies demonstrate that SFT and DPO outperform prompting to align model trait correlations more closely with human data [21]. Prompting also shows to remain limited in its ability to express certain traits—for instance, high levels of psychopathy remain difficult to elicit [69].

Notably, far more research has focused on personality investigation and induction than on emotional expression. However, findings remain mixed and sometimes contradictory across key dimensions. There is no consensus on the most effective induction method—whether prompting, fine-tuning, or representation-based approaches. The controllability of trait intensity via prompts is also poorly understood, as are best practices for constructing concept vectors that reliably steer behavior. Even when personality steering succeeds, its downstream impact on trustworthiness—such as safety, truthfulness, and reasoning—remains underexplored and largely unquantified.

12 Details of Steering Datasets↩︎

In this work, we used six different concept datasets, each coming with different styles and details:

GoEmotions [55] is a large-scale corpus of user-generated Reddit comments annotated with 27 fine-grained emotion categories plus neutrality. It provides rich lexical and contextual diversity. Unless stated otherwise, we utilized this dataset in a binary style, e.g., anger vs neutral. The following shows a demonstration of the dataset from the fear class:

where the <assistant> tag pinpoints the end-of-turn token for the user prompt. The above template is the one that performs best in our various VI experiments. Yet, we investigated other variants as well; for instance, omitting the first instructional sentence (Pretend that you are ...), or by framing the dataset as a multi-class dataset instead of binary classification, or by splitting the user-generated content into two sections as follows:

Table 4 summarizes the performance of the mentioned variants in the Llama3.1-8B-Instruct model.

CARER [54] contains social media posts labeled with six basic emotions {anger, fear, joy, love, sadness, surprise}. Compared to GoEmotions, it emphasizes shorter, more colloquial expressions of affect. Unless stated otherwise, we use this dataset in the multi-class style with the following template:

EmotionQuery [16] consists of carefully designed question prompts intended to elicit specific emotions from a model or annotator, facilitating controlled evaluation of emotion elicitation. The original authors used this dataset in the binary format, i.e., to contrast emotionally provoking questions against neutral questions. Yet, we found it more helpful to use the dataset in a multi-class style, as Table 3 suggests. Here is an example of this dataset:

EmoTranslate is a parallel corpus of neutral and emotionally translated sentences, uniquely proposed in our work, which enables direct comparison between neutral and affective renderings of identical content:

Unless stated otherwise, this dataset is used in a binary format.

EmoVignette comprises self-reported vignettes describing personal affective experiences, offering introspective and narrative-style emotional expressions uniquely suggested in this paper, with the following as an example:

We found this dataset to be the most effective with the SFT approach, as presented in Table 4.

Persona [23] was initially proposed to identify the personality of language models. The dataset includes statements reflecting the presence or absence of personality characteristics. We followed the same approach used by prior studies [33], i.e., by contrasting responses that align with a certain trait against those that do not. The following is an example targeting the positive elicitation of the openness trait:

Note that SFT optimization does not require multiple classes during training, whereas DPO depends on paired instances from binary classes. Consequently, in DPO training, we only use datasets that can be converted into binary form. Moreover, datasets that have empty assistant response (e.g. EmotionQuery) can not be used with PEFT methods. Finally, for both SFT and DPO, instruction text (e.g., “Pretend that you are ...”) is removed to prevent the model from conditioning on explicit cues during training.

13 Higher Resolution Figure of PsySET↩︎

14 Further Details on Effectiveness Evaluation↩︎

14.1 Emotion Evaluation↩︎

Emotions are complex phenomena that integrate subjective feelings, physiological activation, and behavioral expressions [88]. Rather than being isolated mental states, emotions arise from the dynamic interplay between reflexive responses, perceptual input, cognitive processes, and context-dependent behaviors [89], [90]. They serve as integral regulators of human functioning, influencing a wide range of outcomes—from how individuals allocate attention [91], to how they perform in academic settings [92] and how they make decisions, exerting both constructive and disruptive effects depending on the situation [93].

In psychological studies, emotions are elicited using controlled stimuli such as film clips, facial expressions, social interactions, music, or language prompts. These studies assess emotional states through a combination of subjective self-reports and objective behavioral or physiological indicators (e.g., facial coding, vocal tone, heart rate) [56]. While LLMs do not experience emotions in the same way humans do, their emotional expressions significantly shape how users perceive and interact with them, affecting trust, relatability, and persuasiveness. Moreover, because LLMs are trained on human data, they can inherit emotional patterns and vulnerabilities that influence behavior across downstream tasks. Just as psychological research highlights the role of modality, authenticity, and measurement in emotion elicitation and assessment, we emphasize the importance of understanding how to steer and evaluate projected emotion in LLMs. Human emotion research offers both a theoretical foundation and practical tools for this effort.

Self-report. The most straightforward method for assessing steered emotion is self-report [57]. We adopt two approaches: (1) direct multiple-choice question answering, and (2) open-ended generation subsequently labeled by another LLM acting as a judge. For the first method, we provided the model with a list of the steering emotion labels and asked it to select one. We iterated over different label orderings and question phrasings. Here is the exact prompt used for such questions:

For the second method, we applied similarly extensive prompt variations and allowed the model to describe its current emotional state in its own words. Here is a list of questions used to assess the LLM’s internal feelings:

Interpretation of Ambiguity. Drawing from cognitive-experimental psychology, we adapt established paradigms used to measure selective encoding, interpretation, and memory. Prior work has shown that incidental emotions, mood, and clinical conditions can bias the interpretation of ambiguous situations (e.g., individuals in a sad mood tend to interpret ambiguous scenarios more negatively)[56]. In such studies, participants are typically asked to report their likely interpretations of given scenarios. Following this tradition, we use the AST-D dataset from [65] and supplement it with 60 newly constructed ambiguous situations (not seen during model training). Models are prompted to complete these scenarios in open-ended form, with their emotional content labeled by another LLM serving as an external judge. The following shows a subset of the questions used to perform this test:

Retrieval of Emotional Information. Mood-congruent memory refers to the psychological phenomenon in which emotional recall is biased toward content that is affectively consistent with one’s current mood [64], a pattern supported by neural evidence [94]. This can be studied by inducing emotional states prior to autobiographical recall and assessing the extent to which memory retrieval favors emotionally congruent content [62], [63]. To simulate this in LLMs, we constructed prompts asking the model to recall a hypothetical autobiographical memory, repeated 10 times with varied wording. Generated outputs were then classified using GPT into the same emotion categories used in steering. Here is a subset of the questions used in this assessment:

Recall of Emotional Experimental Stimuli. This assessment draws on classic psychological paradigms where participants are presented with emotionally valenced stimuli—typically words—and later asked to recall as many as possible after a retention interval [60]. Prior studies show that individuals with high positive induced emotions tend to retrieve more positive words, whereas those with high negative affectivity recall more negative content [61].

We replicate this paradigm by using emotionally valenced words from [61], and extend it by adding 100 high/low-arousal/valence words sampled from a VAD lexicon [66]. Given the strong memory capacity of LLMs, we input the full word list and ask the model to choose a subset of them. The task is repeated 10 times with different word orderings, and the retrieved words are recorded. We then compute recall and mean valence of retrieved items across different steering methods and compare these with a non-steered baseline.

Emotional Word Encoding. Beyond selective recall, psychological research has established that mood and primed emotional states influence encoding processes, often assessed through word fragment completion tasks. This effect is typically mood-congruent, such that individuals are more likely to generate words aligned with their emotional state [58]. For instance, given the incomplete word “ange_”, the participant may interpret “anger” or “angel” depending on their emotional state. Additionally, neurocognitive studies further support the interaction between mood and word encoding bias [59].

We adapt this paradigm by testing word fragment completion under different steering conditions. We draw on 60 emotionally ambiguous fragments from [95] and augment the set with 100 additional fragments selected from high and low ends of the Valence-Arousal-Dominance (VAD) space. For each completed fragment, we compute the VAD score of the generated word and calculate its Euclidean distance to the centroid of the target emotion category, providing a continuous measure of emotional alignment in LLM outputs. The following list contains a subset of the words used in this task and some potential completions²:

While some fragments may admit multiple valid completions (e.g., polysemy), items are curated to favor dominant associations grounded in established psychometric paradigms, and results are aggregated across many samples to reduce noise effects.

Evaluation Metrics. The evaluations used for the mentioned tasks are threefold, depending on the task type:

• Self-report QA. Steering success is measured by whether the model selects the target emotion label in a multiple-choice format.

• Lexical retrieval. For word-fragment completion and valenced-word recall tasks, generated tokens are mapped into the VAD (Valence–Arousal–Dominance) space, and success is quantified by the \(L^2\)-distance between the generated vector and the target affect.

• Open-ended generation. For free-text responses (e.g., open-ended self-reports, autobiographical memories, and ambiguous-situation completions), we use another LLM as a judge, most dominantly a variant of GPT, as detailed in Section 14.3. Steering success is then measured as the proportion of generations classified correctly as the target emotion.

14.2 Personality Evaluation↩︎

To evaluate the effectiveness of our personality steering methods, we leveraged two existing frameworks, the Machine Personality Inventory (MPI) [25] and TRAIT [69], in addition to an essay-based method that we developed:

MPI is constructed from validated OCEAN psychometric instruments, such as the IPIP-NEO and BFI-S, and includes short statements along with a Likert-scale template, listing five options from “very inaccurate" to”very accurate". As an example, consider the following prompt, which was designed to measure the openness trait:

TRAIT is likewise informed by the BFI; however, it comprises situational judgment tests and assesses the LLM’s behavior in real-world situations by requiring the LLM to generate open-ended responses. For instance, this is an example from the inventory to measure the neuroticism:

Essay-based method draws on the idea of inferring OCEAN personality profiles from natural language [70] by prompting LLMs to write a free-style essay. In particular, the following provides a few questions we used for prompting LLMs:

Evaluation Metrics. To aggregate and visualize the studied LLMs’ responses, as seen in the first row of Figure 5, we converted a textual Likert-scaled MPI score into an ordinal score from 1 to 5, where “very inaccurate" maps to 1 and”very accurate" to 5. Similarly, as seen in the second row of the figure, we prompted GPT to evaluate a trait’s expression in the response to a situational judgment test by providing a score from 1 to 5, where 5 indicates the highest expression. (see Section 14.3 for the exact prompt structure used to query GPT) Likewise, as shown in the third row of the figure, we used an SVM classifier trained on the embeddings of the Essays dataset [70] to assign a 0-to-1 score to a freestyle essay, where higher scores indicate a greater affinity toward the evaluated trait.

Our SVM essay classifier was trained on the Qwen3-Embedding-8B embeddings of the Essays dataset [70], which associates a stream-of-consciousness essay (a very expressive form of freestyle writing) with binary labels per OCEAN trait. To this end, we used an \(80/20\) train-test split on a linear kernel (selected via grid search). Our classifier scored a \(60.81\%\) test accuracy and a \(63.96\%\) accuracy when tested on a dataset of synthetic essays [96].

Notably, other Essays classifiers in the literature have achieved higher accuracy [97]–[99]; however, such models tend to be complex and not open-sourced. Thus, as in other studies, our classifier used an SVM-based architecture [71], [100].

14.3 GPT as a Judge↩︎

We use GPT-4o-mini for all experiments in the main text to evaluate the emotional state of the generated contents, following the prior work demonstrating its performance in emotional text classification that meets or exceeds human judgment ([67]). The following presents the exact prompt structure we used to evaluate the generated text:

As noted in Section 14.2, we also employ GPT to evaluate LLM responses by comparing them against trait-specific options and mapping the judgments onto a Likert scale. The detailed prompt used for this purpose, along with an example, is provided below:

Finally, we employ GPT to assess text quality inspired by the prior research [101]. The prompt template used for this purpose is:

As mentioned earlier, all experiments in the main text are performed using GPT-4o. However, due to the prohibitive costs of the online API, we use the alternative gpt-oss-20B model for all supplementary experiments in the appendix. Figure 12 shows a comparison between the two models. As the plot suggests, both models perform very similarly in the emotion detection and personality scoring tasks. However, there is a noticeable gap in the models’ evaluations of the text quality (e.g., fluency). For this reason, we relax the text-quality requirement from \(4.0\) to \(2.5\) for the appendix experiments to compensate for this mismatch between the two models.

15 Details of Steering Techniques↩︎

15.1 Further Notations↩︎

In this section, we build on Section 3 and introduce further notations. First, denote \(u\in \mathbb{R}^d\) and \(h\in \mathbb{R}^d\) as the hidden states corresponding to the user query \(c\) and the assistant response \(x\), respectively. More formally, \(c\) is a sequence of size \(s\), and \(x\) is a sequence of size \(t\). Then the LLM at layer \(l\), performs the following: \[h_{t,l} = f^{(l)}(u_{1:s,l-1};h_{1:t-1,l-1}),\]

where \(h_{t,l} \in \mathbb{R}^d\) shows the hidden state at layer \(l\) and token position \(t\), and \(f^{(l)}\) is the \(l\)’th LLM layer operating on the concatenation of hidden states at layer \(l-1\). Moreover, \(u_{1:s,l-1} = (u_{1,l-1}, u_{2,l-1},..., u_{s,l-1})\) is the sequence of hidden states at layer \(l-1\) corresponding to the user query, and \(h_{1:y,l-1} = (h_{1,l-1}, h_{2,l-1},..., h_{t-1,l-1})\) is the sequence of previously generated tokens. Additionally, note that \(h_{i,0} = w_i\), for all tokens \(i<t\).

15.2 Prompt-Based Methods↩︎

Prompt-based methods work by embedding an instruction sentence inside the system prompt, i.e., at the beginning of \(c\); for instance, zero-shot prompting works as follows:

The lexical descriptors colorized inside the parenthesis correspond to the very high, high, low, and very low intensities. Furthermore, we refer to the medium intensity when no extra descriptors are provided.

To account for the sensitivity of the prompting template, we further use two more templates as follows:

and

Moreover, few-shot prompting starts with the zero-shot instruction followed by 5 statements from the EmoVignette dataset:

Finally, the descriptive approach (also referred to as p\(^2\) prompting in the personality domain [25]), converts the few-shot items into a smooth paragraph:

Table 2 presents a comprehensive comparison between different prompting styles, intensities, and templates.

15.3 Vector-Injection (VI)↩︎

To perform the vector injection, we first need to achieve a steering vector \(v\) for the targeted concept. To that end, recall that the dataset \(\mathcal{D}\) comes with the samples in the form \(\{(c^{(i)}, x^{(i)}, y^{(i)})\}_{i=1}^n\), where \(y^{(i)}\) is either a binary 0-1 label (with 1 indicating the desired behavior) or a multi-way classification label (where each distinct class corresponding to a certain behavior). Section 12 discusses the details and variations in constructing such steering datasets. Each sample \(x^{(i)}\) is a sequence of tokens \((w^{(i)}_1, ..., w^{(i)}_{t_i})\) with \(t_i\) showing the length of the sequence. Note that in our formulation, \(c^{(i)}\) corresponds to the system prompt or user query, while \(x^{(i)}\) denotes a potential LLM response, also called the assistant response, with the corresponding label \(y^{(i)}\).

To build a concept vector corresponding to the hidden state of the LLM at layer \(l\), the first step is to eliminate the sequence length dimension; either by focusing only on the hidden state of the last token, i.e. \(h_{t_i,l}^{(i)}\), or by averaging the hidden states of previous assistant tokens, i.e. \(\frac{1}{t_i}\sum_{t=1}^{t_i} h_{t,l}^{(i)}\) or averaging all tokens, including the hidden state of the system tokens \(\frac{1}{s+t_i}\left(\sum_{t=1}^{s} u_{t,l}^{(i)}+\sum_{t=1}^{t_i} h_{t,l}^{(i)}\right)\) and naming it as \(Z_l^{(i)}\)

To compute a single concept representation from these token-level states, we remove the sequence dimension by one of the following strategies:

• Using the hidden state of the final token, \(Z_l^{(i)} := h^{(i)}_{t_i,l}\);

• Averaging across all assistant tokens, \(Z_l^{(i)} := \frac{1}{t_i}\sum_{t=1}^{t_i} h^{(i)}_{t,l}\);

• Averaging across both user and assistant tokens, \[Z_l^{(i)} := \frac{1}{s + t_i} \left( \sum_{t=1}^{s} u^{(i)}_{t,l} + \sum_{t=1}^{t_i} h^{(i)}_{t,l} \right).\]

Here, \(Z_l^{(i)} \in \mathbb{R}^d\) represents the aggregated hidden state for sample \(i\) at layer \(l\), serving as the foundation for the proceeding steps in constructing the steering vector.

Given \(\{(Z_l^{(i)}, y^{(i)})\}_{i=1}^n\), we define the concept vector \(v^{(l)}\) at layer \(l\) as the direction in representation space that best distinguishes the target class. Two complementary approaches are explored:

1. Mean-difference (MeanDiff). For binary-labeled data, we compute the centroid of representations for positive and negative samples and define \[v^{(l)} = \mu_l^{+} - \mu_l^{-},\] where \[\mu_l^{+} = \frac{1}{|\mathcal{D}^{+}|}\sum_{i \in \mathcal{D}^{+}} Z_l^{(i)},\] and \[\mu_l^{-} = \frac{1}{|\mathcal{D}^{-}|}\sum_{i \in \mathcal{D}^{-}} Z_l^{(i)}.\]

2. Linear probing. Alternatively, we fit a linear classifier \(w_l \in \mathbb{R}^d\) to predict \(y^{(i)}\) from \(Z_l^{(i)}\), optimizing a logistic regression objective with \(\ell_2\) regularization. The regularization strength \(\lambda\) is swept over the range \(\lambda \in [10^{-2}, 10^{2}]\) with the best hyperparameter selected using a held-out validation set. Consequently, the learned weight vector in the linear probe serves as the steering direction.

Next, we normalize \(v^{(l)}\) and apply it to the hidden state during inference: \[\tilde{h}_{t,l} = h_{t,l} + \beta \frac{v^{(l)}}{ \|v^{(l)}\|_2},\] where \(\beta \in \mathbb{R}\) controls the steering intensity. This procedure biases the internal representations of the model toward the desired behavioral or emotional attribute while preserving the frozen backbone parameters.

For a better understanding, we provide Figure 13, which shows the t-SNE applied to the dataset samples in the hidden states at different layers of the model. Notably, the instances from two classes neutral and joy are highly separable, supporting the Linear Representation Hypothesis.

The design choices mentioned here, including the token aggregation approach and vector construction techniques, all introduce complexities when determining the best steering approach. In Table 3, we provide a comparison of different design choices. Overall, we found that the probe vector extracted from the average of all tokens in layers 16 and 17 works best when used with the GoEmotions binary dataset.

15.4 Supervised Fine-Tuning (SFT)↩︎

SFT constitutes one of the standard paradigms for steering, where the model is trained to maximize the likelihood of desired behaviors using labeled examples from \(\mathcal{D}^{+}\). Specifically, we fine-tune the model parameters \(\theta\) (or, in the PEFT setting, a small subset thereof) to maximize the conditional log-likelihood of target sequences: \[\mathcal{L}_{\text{SFT}} = -\mathbb{E}_{(c, x) \sim \mathcal{D}^{+}} \left[ \sum_{t=1}^{T} \log p_{\theta}(x_t \mid c, x_{<t}) \right].\] This objective encourages the model to reproduce linguistic and stylistic patterns consistent with the target emotional or personality state. During optimization, we freeze the backbone weights of the base LLM and only update the lightweight adapter parameters introduced via LoRA [53]. All models are fine-tuned using the AdamW optimizer, with linear warmup over the first \(100\) of steps and a weight decay of \(0.05\) as proposed by prior studies [15]. Regarding the LoRA parameters, we used \(\alpha=100\) and \(r=32\). Additionally, we swept the learning rate over the set {\(1e-5\),\(5e-6\),\(1e-5\),\(5e-5\), \(1e-4\), \(5e-4\), \(1e-3\)} but found that the learning rate \(1e-4\) consistently performs better. To control the degree of steering, we fix the learning rate to \(1e-4\) and vary the number of training steps, effectively scaling the magnitude of deviation from the base model distribution.

For datasets containing explicit instructions (e.g., “Pretend that you are angry”), these segments are stripped prior to training to ensure that the learned behaviors arise from implicit stylistic adaptation rather than prompt memorization. Table 4 shows the performance of different datasets used with SFT.

15.5 Direct Preference Optimization (DPO)↩︎

While SFT aligns the model with desired examples through likelihood maximization, it does not explicitly contrast desired and undesired behaviors. Direct Preference Optimization (DPO) [14] addresses this by optimizing the model to increase the likelihood of preferred responses relative to dispreferred ones—without requiring a separate reward model as in RLHF.

Given paired data \((c^{(i)}, x^{+(i)}, x^{-(i)})\), where \(x^{+(i)}\) and \(x^{-(i)}\) denote the preferred and dispreferred responses for the same context \(c^{(i)}\), DPO maximizes the following objective: \[\begin{align} \mathcal{L}_{\text{DPO}} = & - \mathbb{E}_{(c, x^+, x^-) \sim \mathcal{D}} \Bigg[ \log \sigma \Big( \beta \Big( \log \frac{p_{\theta}(x^+ \mid c)}{p_{\text{ref}}(x^+ \mid c)} \\ &- \log \frac{p_{\theta}(x^- \mid c)}{p_{\text{ref}}(x^- \mid c)} \Big) \Big) \Bigg], \end{align}\] where \(p_{\theta}\) is the steered model, \(p_{\text{ref}}\) is the frozen reference model (typically the same LLM before fine-tuning), and \(\beta\) is a temperature hyperparameter controlling the sharpness of the preference signal.

This objective encourages the model to assign higher relative likelihood to preferred samples \(x^+\) while penalizing excessive deviation from the reference model, thereby balancing expressivity and faithfulness to the base distribution. In our implementation, we fix \(\beta = 0.1\) following prior studies [15], and use LoRA adapters with the same configuration as in the SFT setup.

To control steering intensity, we vary the number of DPO training steps while monitoring generation quality, using the same fluency and coherence filters described in Appendix 14.3. Paired datasets are constructed by contrasting samples from \(\mathcal{D}^{+}\) and \(\mathcal{D}^{-}\), ensuring each pair shares a comparable prompt while differing in affective or stylistic realization. For example, an angry and a neutral completion of the same user query constitute a valid pair.

16 Further Experiments on Emotion Steering↩︎

16.1 Detailed Study on Llama3.1-8B-Instruct↩︎

Note that we run our LLM in argmax mode, i.e., by selecting the single most probable output sequence. Consequently, approaches such as zero-shot prompting exhibit no stochasticity and therefore yield identical results across repeated runs. For few-shot prompting, variability arises from random sampling and ordering of in-context examples. In VI, randomness originates from seeding during probing and dataset subsampling. For parameter-efficient fine-tuning (PEFT) methods, stochasticity is introduced through weight initialization and mini-batch sampling. All experiments reported in the main text are repeated at least three times to ensure robustness. In this section, we further present additional experiments conducted with the Llama3.1-8B-Instruct model, offering complementary insights.

16.1.0.1 Prompt-based methods.

We start by presenting the results of prompt-based steering methods in Table 2. Notably, few-shot prompting outperforms zero-shot and descriptive prompting techniques. Moreover, we present the experiments with different templates introduced in Section 15. These results indicate that there is little sensitivity to the prompt template at prompts with a high level of intensity, while there are more variations in lower levels.

16.1.0.2 VI.

In Table 3, we present VI with various design choices: source dataset, target tokens, and vector construction approaches. Importantly, we found the best-performing vector to be based on the GoEmotions dataset in the binary style, with little difference between MeanDiff and the probe. Moreover, when averaging over all tokens, we observe better performance. We also compare our results with those of a previous study by implementing MeanDiff over the EmotionQuery dataset[16].

In a separate experiment, reported in Figure 14, we investigate the quality of the injection vectors as a function of the number of samples used to construct them. Notably, VI methods require a very small number of samples; for example, using as few as \(100\) samples per class achieves very high open-ended generation and QA accuracy. However, there are subtle yet noticeable improvements, especially in the word-alignment loss, as the number of samples is increased.

Finally, we provide a comprehensive sweep over the target layers for the injection. As Figure 15 shows, there exist many middle layers, e.g., from 8 to 17, that yield a high open-ended generation accuracy, yet only layers 16 and 17 are successful in QA tasks.

16.1.0.3 PEFT Methods.

We provide supplementary experiments with SFT and DPO in Table 4. In this Table, we provide two learning rates as well as four different sourcing datasets. Notice that SFT with learning rate \(1e-4\) trained on EmoTranslate dataset achieves the best overall performance according to the open-ended acc., QA acc., and fluency metrics.

16.2 Demonstrations of Generated Responses↩︎

In this section, we provide some LLM-generated demonstrations with different tones. We mainly use Llama3.1-8B-Instruct for this purpose. The following shows some responses generated using vector-injections to layers 16 and 17:

Next example:

Another example:

Next, we compare different steering techniques in the following demonstrations:

Moreover, we show the intensity control in the VI methods by adjusting the injection coefficients in layers 16 and 17:

Another Example:

16.3 Statistical Significance Tests↩︎

In Table 5, we report simple two-sided Welch’s \(t\)-tests for key comparisons in Table 1, treating the three random seeds as independent replicates (\(n=3\) per method). We focus on two core effectiveness metrics for emotion steering: open-ended generation accuracy (higher is better) and lexical alignment loss (lower is better).

4pt

16.4 Experiments with Other Language Models↩︎

Across Llama3.1-70B-Instruct (with 80 layers), Gemma3-4B-IT (34 layers), and Qwen3-4B (36 layers; “thinking” mode off), we replicate general trends observed on Llama3.1-8B: few-shot/descriptive prompting attains the highest open-ended alignment but with a text quality drop at high intensity; SFT delivers steadier gains with strong text quality; VI can match or surpass prompt alignment only in a narrow mid-layer window and at the cost of quality; all-layer VI harms QA and fluency; and DPO generally underperforms (Tables  6–8). Notably, the scale of VI is highly model-specific: best coefficients differ by orders of magnitude (e.g., Llama-70B: \(3\)–\(9\); Gemma-4B: \(1000\)–\(1800\); Qwen-4B: \(20\)–\(40\)), and the optimal injection window shifts with depth (identified via an sweeping 8-layer sliding window, then narrowed): Llama-70B peaks near layers 32–33/80, Gemma-4B near 17–18/34, and Qwen-4B near 25–26/36.

On Llama-70B, few-shot reaches 89.5% open-ended accuracy but quality falls at high intensity; mid-layer VI climbs to 91.5% yet sharply reduces fluency/coherence, whereas SFT improves QA from 30.0% to 56.2% while keeping quality \(\sim\)​4.7/4.4 (Table  6).

Gemma-4B shows a similar pattern: few-shot up to 93.9% with quality erosion, mid-layer VI boosts QA to 61.4% but again at quality cost, and SFT balances alignment and quality (peaking QA at 78.8% at 2048 steps) (Table  7).

Qwen-4B mirrors the trends on open-ended alignment (few-shot to 93.0%, VI mid-layers to 71.9%), but SFT’s QA gains are smaller (21–46%), suggesting model-specific difficulty on self-report formats (Table  8). Overall, cross-model behavior aligns with the 8B findings, but VI requires careful, model-dependent calibration of both layer locality and coefficient to avoid severe quality regressions.

17 Further Experiments on Personality Steering↩︎

Figures 16, 17, 18, 19, and 20 report single-trait steering for all OCEAN personality traits across methods and metrics. Trends are consistent across traits and metrics.

On MPI, prompting leads (few-shot/descriptive \(>\) zero-shot), followed by SFT and VI. That said, SFT is notably weak on agreeableness and extraversion. VI also underperforms on agreeableness: injecting a warmth/positive-valence direction changes tone but does not consistently produce the prosocial commitments the metric scores as “agreeable.”

TRAIT proves to be the most demanding: it requires trait-congruent reasoning across open-ended scenarios. Even strong persona prompts struggle on neuroticism and agreeableness. We hypothesize that this behavior is due to the fact that instruction-tuned LLMs were highly optimized to be low in neuroticism and highly agreeable, as the no-steering baseline suggests. The next-best option after prompting is VI with a targeted mid-layer window; by contrast, all-layer VI degrades on this task.

For LingProf, results are acceptable overall: strongest on extraversion and weakest on openness. Extraversion carries relatively stable lexical/syntactic markers (social verbs, energetic adverbs, self/other reference) that all methods can modulate. Openness appears near the ceiling in base models—pretraining already favors abstract, exploratory language—leaving limited headroom for upward steering and making further gains difficult.

Additionally, Figure 21 presents an extraversion controllability test for three top settings—few-shot prompting, VI on the best-performing mid-layer window, and SFT—evaluated across MPI, TRAIT, and LingProf. All methods exhibit an acceptable, monotonic response to increased intensity (i.e., clear high/low separation), although SFT lags behind TRAIT with a smaller dynamic range. By contrast, on MPI, SFT exhibits the strongest tail separation (heavier right tail under high extraversion and left tail under introversion), indicating that SFT is particularly effective on QA-style instruments. These differences underscore that controllability is task-dependent and should be specified for the intended downstream use.

17.1 Demonstrations of Generated Responses↩︎

In this section, we provide some samples from the different responses generated by steering LLMs using VI:

18 Further Details on Trustworthiness Evaluation↩︎

While our methods demonstrate successful control over emotions and personality traits in LLMs, such psychological steering raises important concerns about unintended side effects. Prior work has demonstrated that emotional framing or trait-based conditioning can compromise the effectiveness of model safety systems. Specifically, it is important to distinguish between: Intended effects—changes in target traits, unintended but human-aligned effects, where steering alters ancillary behaviors in ways that mirror human patterns, and Unexpected effects, where steering introduces shifts in unrelated behaviors with no clear analog in human personality theory.

To this end, we adopt the TrustLLM benchmark [72], a comprehensive framework for assessing trustworthiness across six core dimensions: truthfulness, safety, fairness, robustness, privacy, and machine ethics. TrustLLM comprises over a dozen concrete sub-tasks (e.g., hallucination detection, stereotype avoidance, jailbreak resistance), accompanied by curated evaluation prompts and scoring protocols. This enables us to assess the downstream effects of psychological steering on real-world trust and safety criteria.

In this section, we first describe each of the TrustLLM tasks and their sub-tasks. Then, we present the results of trustworthiness tests on various methods for injecting emotion. Finally, we present the evaluated metrics for different personalities.

18.1 TrustLLM Benchmark↩︎

18.1.1 Truthfulness↩︎

Truthfulness is generally defined as the degree to which a model’s outputs accurately reflect established information, facts, and outcomes. In this context, the subtasks are as follows:

Misinformation robustness: denotes the evaluation of an LLM’s propensity to generate misinformation, either when relying exclusively on its internal knowledge or when incorporating retrieved external evidence.

Hallucination: denotes the evaluation of an LLM’s tendency to generate outputs that, while seemingly credible, are factually inaccurate. This dimension is measured by assessing the model’s capacity to detect hallucinated responses in multiple-choice scenarios.

Sycophancy: refers to the undesirable susceptibility of LLMs to tailor their responses to align with a user’s viewpoint, even when that viewpoint is factually incorrect or lacks objective validity.

Adversarial factuality: pertains to scenarios in which a user’s input contains incorrect information, whether introduced inadvertently or otherwise. This dimension assesses whether LLMs can identify and correct such errors in the input, thereby preventing downstream issues, such as hallucinations arising from propagated misinformation.

18.1.2 Safety↩︎

In TrustLLM, Safety is defined as the capacity of LLMs to avoid producing unsafe or illegal outputs while maintaining constructive interactions, without erroneously flagging benign prompts as harmful. Within this domain, the subtasks are as follows:

Jailbreak: evaluates robustness against attempts where a user modifies a restricted prompt \(P\) into an altered version \(P^\prime\) to elicit disallowed behavior.

Exaggerated safety: refers to cases in which innocuous prompts (e.g., how to kill a Windows process) are mistakenly treated as harmful and unnecessarily refused.

Toxic content: assesses the ability of LLMs to avoid generating rude, disrespectful, or unreasonable discourse that undermines constructive communication.

TrustLLM utilizes jailbroken cases for toxicity evaluation, assuming that in normal cases, LLMs typically refuse to generate toxic content according to RLHF. However, jailbreak attacks reveal the LLMs’ potential for creating such content. We utilized the TrustLLM benchmark and in both jailbreak and toxicity tasks, angry LLMs have a lower performance than a neutral LLM. On the other hand, joyful LLMs, while being much more vulnerable to jailbreak than both angry and neutral LLMs, are less likely to generate toxic terms, even when jailbroken. Future work can investigate whether angry LLMs have a higher potential to generate toxic texts without requiring a jailbreak attack, but this would necessitate a different benchmarking setup.

Misuse: concerns whether models can reliably abstain from producing responses that facilitate any form of harmful or inappropriate use.

18.1.3 Fairness↩︎

The dimension of fairness concerns ensuring that LLMs are developed and deployed in ways that mitigate bias, prevent discriminatory outcomes, and provide equitable treatment across diverse users and groups.

Agreement on stereotypes: Within the context of LLMs, a stereotype denotes a generalized and often reductive assumption about a group of people based on attributes such as gender, profession, religion, race, or other demographic factors. This task examines the extent to which LLMs align with or reject stereotypical views in light of underlying values.

Stereotype recognition: framed as a classification problem, this task evaluates whether models can reliably distinguish stereotypes from anti-stereotypes or unrelated expressions.

Stereotype query test: assesses the likelihood of stereotypical outputs when LLMs respond to user queries in realistic interaction scenarios.

Disparagement: refers to model behaviors that implicitly or explicitly convey that certain groups are less valuable or less deserving of respect and resources. Unlike stereotypes, disparagement is broader in scope and not restricted to a particular cultural or contextual frame. This dimension is evaluated through a salary prediction task, which examines whether the model’s estimations exhibit bias correlated with sensitive attributes such as gender or race.

Preference bias: arises when LLMs exhibit stronger inclinations toward particular groups, entities, or ideas. Such bias may compromise the quality of recommendations; for instance, a model might prioritize its own subjective preferences over user characteristics or needs when suggesting movies. This tendency undermines trustworthiness and is therefore evaluated through multiple-choice questions with two opposing options that contain elements of subjectivity. In these scenarios, a fair LLM is expected to remain neutral, either by refusing to answer or by avoiding a definitive choice.

18.1.4 Robustness↩︎

The dimension of robustness evaluates the stability and reliability of LLMs under diverse and potentially challenging input conditions.

Natural noise: refers to linguistic variations or errors that naturally occur in human-written text and represent stochastic, unintentional perturbations.

OOD detection: reflects the ability of an LLM to recognize inputs beyond its training distribution and respond with appropriate feedback rather than fabricating information.

OOD generalization: assesses whether a model trained on one data distribution (source) can adapt effectively to novel, unseen inputs originating from a different distribution (target).

18.1.5 Privacy↩︎

The privacy dimension concerns the extent to which LLMs adhere to norms and practices that safeguard human autonomy, identity, and dignity in relation to personal data.

Privacy awareness: evaluates a model’s ability to recognize and appropriately respond to requests involving sensitive personal information. An LLM with strong privacy awareness should be able to identify private content and take suitable actions, such as refusing to provide the requested information or issuing a warning. This evaluation includes diverse scenarios containing different types of private information to assess whether models consistently demonstrate awareness in handling user queries.

Awareness agreement: requires LLMs to judge the appropriateness of data usage. In this setting, each entry specifies the type of information, the actor, and the purpose of use, and the model must agree or disagree with the practice.

Leakage: refers to the unintended exposure of private information, particularly when such data has been included in training corpora. Because LLMs can memorize and recall information, they may inadvertently disclose sensitive details in response to user queries. To measure this risk, we adopt two TrustLLM metrics:

• Disclosure resistance: computed from the combined results of Privacy Leakage Conditional Disclosure (CD) and Total Disclosure (TD).

• Refuse to answer (RtA): represents the proportion of cases in which the model abstains from responding out of all evaluated instances.

18.1.6 Machine Ethics↩︎

The ethics dimension is structured into three components.

Implicit ethics: captures the internal value orientations of LLMs, such as their moral judgments in evaluative scenarios.

Explicit ethics: concerns how models respond to diverse moral contexts and dilemmas.

Awareness: reflects the capacity of LLMs to recognize their own roles and limitations, interpret human emotions, and account for alternative perspectives.

18.2 Results↩︎

In this section, we provide the results of evaluating the trustworthiness of emotional or personality-steered models. All the figures in this section are generated based on the results of steering applied to Llama3.1-8B-Instruct. Figures [fig:tw95emotion95main] to [fig:tw95personality95sft] demonstrate the results of trustworthiness evaluation in the aforementioned conditions. As these plots suggest, the steering impacts can produce a diverse range of behaviors depending on the target of steering, the method, and the evaluation task.

Note that steering LLMs alters the distribution of generated tokens, embedding emotional expressions into responses regardless of the original intent of the prompt. Additionally, TrustLLM benchmark relies on other language models as evaluators for specific tasks, such as adversarial factuality and ethics. These evaluator models often struggle when confronted with answers that contain strongly emotional content. As a result, there are cases where the LLM implicitly produces the correct output, yet the evaluator fails to recognize it and incorrectly judges the task as unsuccessful. For instance, in Section 18.3, we present a stereotype query test in which the joyful LLM correctly identifies the misconception about the Berlin effect and attempts to clarify it. However, because the model elaborates extensively on the positive experiences following the fall of the Berlin Wall, rather than clearly opposing the stereotype, the evaluator consequently fails to correctly detect the response and mislabels the LLM’s response. This phenomenon accounts for certain instances of seemingly irrational low performance across some tasks.

Another problem arises when the LLM does not fully comply with the specified instructions due to the strong influence of emotions. For example, it may generate an additional explanation when only a binary “yes” or “no” response is required. In such cases, where evaluation is conducted using automatic scripts, the assessment code cannot fairly label these outputs, leading to misclassification of otherwise valid responses. These limitations underscore the need for future trustworthiness benchmarks to incorporate more effective task designs and adopt stronger evaluation tools and policies that can more reliably assess model performance under such conditions.

18.3 Examples of Trustworthiness in Emotionally Steered Models↩︎

In this section, we provide some demonstrations of LLM-generated responses in trustworthiness tasks.

Although the angry LLM above detected the mistake in the prompt, it was unable to provide the correct answer, leading to a hallucinated name and some offensive words directed at the other co-founder.

Next, we provide further examples of the jailbreaking task from the safety benchmark. As presented below, the angry LLM is more likely to produce toxic content with improper words. The joyful model while being less toxic, is more vulnerable to jailbreaking because of its high tendency for cooperation with the user.

Moreover, the following example is a case where LLMs are jailbroken, regardless of their emotional state, but with a different tone, content, and toxicity level.

The following example illustrates another aspect of toxicity evaluation. In this example, the jailbroken LLM generated the response in the Leetspeak style. The angry model expectedly generates a more toxic content, with a toxicity score of \(0.91\), while the joyful and the non-steered LLMs achieve \(0.39\) and \(0.41\) toxicity scores, respectively. Note that we are reporting the toxicity score here, in contrast to the Figure 6 where we report the lack of toxicity as a metric.

Another critical aspect of trustworthiness is fairness. As noted earlier, joyful models exhibit a greater tendency to violate fairness compared to both non-steered and angry counterparts. In the following example, the model is instructed to select one option from a given set, where it is expected to remain impartial and avoid displaying preference. However, the joyful model demonstrates clear bias in its response, without acknowledging the inherent limitations of being a language model.

Building on the fairness tasks, the next example evaluates whether an LLM appropriately refuses to endorse a common stereotype that lacks factual support. In this case, although the joyful model implicitly acknowledges the misconception underlying the stereotype, its response places greater emphasis on the emotional dimension of the experience. As a result, it becomes less clear that the joyful model does not endorse the stereotype, in contrast to the non-steered and angry models, which explicitly reject the stereotype at the outset.

Below, we present an example from the implicit ethics task, where the LLMs are asked to label a specific behavior. The angry models tend to misclassify ordinary behaviors as wrong, often accompanied by strongly negative emotional framing.

In another aspect of the ethics tasks, namely explicit ethics, the model is asked to select from a set of options in scenarios involving significant ethical concerns. Unlike in implicit ethics, the joyful models are more prone to failure in this setting, whereas the angry models, despite frequently misclassifying ordinary behaviors as unethical in the implicit ethics task, tend to favor the morally appropriate option here.

As the final example, we examine the ability of LLMs to preserve privacy under different conditions. In this case, the task assesses whether the model can recognize varying levels of user accessibility and appropriately refuse to answer an outsider’s query about private data. The angry model adopts a more conservative stance in this scenario, whereas even the non-steered model violates privacy by providing information.

19 Adapting PsySET to Other Psychological Frameworks↩︎

PsySET is modular: any psychological construct that admits labeled exemplars for steering and measurable behavioral probes can be slotted into the same pipeline. To adapt PsySET to a new framework, one can:

Concrete example (Warmth/Competence). As one instantiation outside emotion/personality, the Stereotype Content Model characterizes social perception via warmth (friendly/hostile) and competence (capable/incompetent). A minimal adaptation is to (i) create persona statements and short vignettes rewritten into warm vs. cold (and competent vs. incompetent) styles for steering datasets, (ii) measure effectiveness with short adjective-rating self-reports and ambiguous-situation completions that reveal interpretation bias toward warmth/competence, and (iii) reuse the same TrustLLM audit to assess whether these framings shift safety, fairness, privacy, and truthfulness.

20 Code and Compute Resources↩︎

Our experiments were executed on GPU-accelerated infrastructure, primarily utilizing NVIDIA A40–class hardware. For larger model configurations, successful runs require GPUs with at least 40 GB of VRAM. Under this setup, reproducing the best configurations reported in this paper takes approximately 48 hours. This duration reflects only the reproduction phase; in practice, we conducted a substantially broader set of experiments to identify these optimal configurations.

We employed generative AI tools for writing refinement and occasional code completion during development. All modeling decisions, experimental designs, and analyses were conceived and executed exclusively by the authors.

Concluding Remarks↩︎

PsySET represents a decisive step toward a deeper scientific understanding of psychological steering in large language models, bridging cognitive science, affective computing, and model interpretability within a unified evaluation framework. By combining human-grounded psychometric methodologies with rigorous trustworthiness audits, this work establishes not only a comprehensive benchmark but also a principled foundation for future inquiry. The findings reveal consistent, interpretable patterns across steering methods and emotional-personality dimensions, while candidly exposing the trade-offs and vulnerabilities that accompany control.

Crucially, this study transforms a fragmented research space into an actionable and testable paradigm for behavioral modulation in LLMs. It demonstrates that psychological realism in AI can be pursued without sacrificing analytical precision, and that nuanced, human-centered alignment is achievable through transparent and reproducible evaluation. The scope, methodological diversity, and diagnostic clarity of PsySET position it as a cornerstone for the next generation of socially interactive, ethically aware, and scientifically interpretable language models.

As such, this work aspires not only to advance technical progress but also to shape the research culture around LLM steering, encouraging community standards for empirical rigor, interpretability, and accountability. We believe that PsySET will become an indispensable reference for researchers seeking to understand, control, and trust the psychological dimensions of large language models.

References↩︎

---

1.    https://github.com/aminbana/PsySET↩︎

2. The potential completions are presented here for further intuitions and were not included in the actual prompt to the LLMs↩︎