Do LLMs have Consistent Values?

A key goal of Large Language Models (LLMs) is to produce agents that will be able to communicate in a “human-like” fashion. However, human communication is characterised by some level of consistency within an individual, as well as variability between individuals. This raises a key question: during a single conversation with an LLM, does the “LLM-persona” resemble a single human? Furthermore, across multiple conversations, can LLMs produce multiple personas that resemble a population of humans? If this is indeed possible, how can such personas be elicited to best resemble psychological characteristics observed in human populations?

This question has only recently begun to be addressed. For example [1] show how probing LLMs with different names leads to variability which in some cases agrees with that of human populations. Here our focus is on understanding whether an LLM in a single conversation can exhibit a psychological characteristic profile that is similar to that of humans. This is a highly challenging question, since it requires analyzing a complete conversation and evaluating whether it conceivably could have been generated by a single individual.

In order to stand on more quantitative ground to evaluating the quality of output relative to human research, we turn to the well established field of value psychology. Namely, we aim to quantify the values that LLM responses are aligned with, and whether these are in agreement with the value hierarchy and structure observed in humans. The question of values in LLMs has rarely been studied, and is of naturally broad interest. As an example of recent work, [2] prompt an LLM with a description of a profile of an individual characterized by a value and check whether generated text is consistent with this description. Our focus is very different, and asks whether an LLM response is in agreement with what we expect human responses to look like given research in the field.

Values are basic motivations that play a foundational role in psychology, influencing perceptions and behaviors across various domains [3, 4], and representing fundamental aspects of human personality [5]. They have been extensively studied across diverse populations, and research has consistently demonstrated their enduring influence over behavior across time and contexts [4].

One prominent framework for studying values, the Theory of Basic Human Values [6], outlines 19 core values, categorized by motivational goals [7]. These values can be simplified into a two-dimensional structure: conservation vs. openness to change, and self-enhancement vs. self-transcendence. The theory describes interrelations among values, suggesting that motivations driving some values are compatible with those driving other values, yet conflict with those underlying yet others. For instance, pursuing independence and creativity (self-direction) aligns with seeking change and variability (stimulation), but conflicts with an emphasis on the status quo (conformity). See Figure 1 for the theorized circle. Despite individual differences, there is a universal hierarchy, where for example caring for close others ranks high, while values related to dominance hold less importance across societies [8]. There is also ample empirical evidence that compatible values do tend to co-occur in humans [9, 10, 11], and are thus a “marker” of a human-like value system.

figurec

Our key question is therefore whether LLM responses demonstrate the same statistical behavior observed in humans: value-ranking and value-correlations. Note that this question is only meaningful when observing a single session of an LLM, where it can exhibit multiple values, because different sessions may be expressing different LLM “personas”. To study this question, we present LLMs with a value questionnaire (the Portrait Value Questionnaire—Revised – PVQ-RR– from [13], a well-established measure of values), and prompt them to answer all the questions in a single session (i.e., in the same context window). We then analyze the provided answers, putting specific emphasis on the correlation between answers in the same session.

We analyze two of the most recent large language models (LLMs): GPT 4 and Gemini Pro¹¹1Our analysis also included GPT-3.5 and Palm2, which produced qualitatively similar results.. Our examination reveals that standard prompting of LLMs does not result in a population of human-like personas. We go on to explore prompting the LLMs with other prompts that provide additional information about the LLM persona. In particular we consider names [1] and persona descriptions. In addition, we consider a novel prompt which we refer to as a “Value Anchor”, which instructs the language model to answer as a person emphasizing a given value. We find that with these prompts, and in particular with the Value Anchor prompt, the overall statistics of the LLM responses (across sessions and within a session) closely mirror those of human subjects. Perhaps most surprising is our finding that the correlation between values agrees with the well known Schwartz circular model for correlations between values.

Our main findings are thus that LLMs cannot be considered as having a consistent, human like, value system that can be studied as representative of the values of the LLM. However, with appropriate prompting, LLMs can exhibit various coherent value personas similar to that of a population of humans. In other words, they answer a set of questions in a way that is consistent with a response from multiple human subjects. We provide information for best prompts and settings for this aim. In addition, we provide two datasets comprising 300 personas each, generated by GPT 4 and Gemini Pro. In conclusion, our results demonstrate the utility of using psychological theory for evaluating consistency of personas generated by LLMs.

Values in LLMs: Our work is based on the Schwartz theory of Personal Values, a highly accepted theory within personality psychology [3]. Values are abstract goals, defining the end states individuals aspire for (e.g., safety, independence), used to direct judgements and behaviors [6, 7]. Individuals typically prioritize their values, so that values stemming from compatible motivations are similarly important, while values stemming from conflicting motivations are prioritized differently. These associations were replicated across hundreds of samples, across the world [14, 15], and make value theory especially useful to identify the coherence of the value profiles created by LLMs. Several studies assumed LLMs can be characterized as operating on the basis of a single set of values, taking an “LLMs as individuals” approach. Fischer and colleagues [2] tested whether ChatGPT could comprehend human values by providing it with value-related prompts and analyzing whether its responses matched the intended value category. A second study [16], compared ChatGPT’s values to those observed in the World Value Survey, while another [17] investigated how temperature influences GPT-3’s responses to the Human Value Scale. Scherrer and colleagues [18] studied responses of LLMs prompts evaluating moral positions, especially in ambiguous settings. A recent study [19], tested the value-like constructs embedded in LLMs and revealed both similarities and differences between LLMs and humans’ values. Kovak et al. [20] challenged those studies by establishing that context starkly influences values expressed by ChatGPT. They found significant variability in ChatGPT’s value expression in response to contextual changes, threatening the notion of stable characteristics of LLMs. Building upon the insights in [20], our study posits that upon providing controlled variability of context, LLMs can elicit a population of multiple personas. In this regard, we aim to further explore the accuracy of LLMs’ mimicking abilities within a controlled experimental framework.
Prompting LLMs: There is extensive research on the prompt design for mimicking individual characteristics in LLM responses [21]. Approaches use specific scenarios [22], questionnaire items [23], simulation of social identities or areas of expertise [24], utilization of titles and surnames representing genders and ethnicities [1], and other demographic information [25]. Additionally, researchers explored the use of designated personas [26], and employed RLHF [27] to guide LLMs to reflect distinct personality traits. Despite this extensive body of work, to our knowledge, no study has directly compared the various prompting techniques to determine which approach yields responses that simulate within-session psychological characteristics of an individual best.
Temperature in LLMs: Adjusting the temperature stands as a common practice for introducing variability in LLM responses [17]. However, consensus is lacking on the optimal temperature setting in simulating psychological characteristics. Some researchers advocate for higher temperatures to boost creativity [24], yet this can also introduce more noise into the data [28]. Conversely, setting the temperature to zero minimizes variability, enhancing replicability [27], albeit posing challenges for variance-dependent analysis [29]. Our framework enables us to explore how temperature adjustments impact the ability of LLMs to simulate human characteristics across multiple datasets.
Evaluating the Quality of Persona Generation in LLMs: The ability of LLMs to mimic and portray human characteristics is a focus of intense research [30, 31]. LLMs can express psychological traits and attributes similar to human individuals [27, 32], and even simulate diverse populations [33, 24]. However, we are only beginning to understand the coherence of these LLM-generated characteristics in mirroring human psychological profiles [1, 20], and how to reliably produce such responses. We are specifically challenged to evaluate the coherence of the resulting psychological profiles. The literature suggested a number of approaches, including an open-ended interview with LLM-generated personas in order to assess the consistency between their intended characters and the responses [34]. In addition, one may apply an additional “judge” LLM in order to check an LLM persona [35]. Finally, [23] assessed coherence with a description used to prompt the LLM. Our study extends upon this line of research by applying well established characteristics of human psychology to investigate the quality of LLM generated personas.

In this section, we introduce the experimental design, models and prompts. The code and data are provided as supplementary files in the submission.

The Value Questionnaire: Our key goal was to assess responses of LLMs to questionnaires used to measure values in human subjects. Specifically, we considered the commonly used 57-item Portrait Value Questionnaire—Revised (PVQ-RR; [13]), developed to measure the 19 values in the Schwartz’s theory. The questionnaire describes fictional individuals and what is important to them. For example: “It is important to him/her to take care of people he/she is close to” (an item measuring benevolence-care values). For each such item, the subject is requested to indicate on a 6-point scale to what degree the persona they form is similar to the person described. Answers are categorical and range from a value of 1 (indicating “not like me at all”) to 6 (indicating “very much like me”). See Appendix for instructions and more example items from the questionnaire.
Models Used: We employed two prominent LLMs, specifically OpenAI’s GPT-4 and Google’s Gemini Pro. Each model was prompted with the five prompts (see Section 3.1), 300 times overall. Half of the runs applied the male-version of the questionnaire, and half the female version. The entire process was conducted twice, once with the temperature parameter set to $0.0$ and once with it set to $0.7$, resulting in the generation of 20 datasets for analysis.

In the current study we analyzed the values exhibited in LLM responses. We used three metrics to estimate the quality of the LLM responses against human responses: value ranking, internal consistency, and value correlations. Our results highlighted the importance of the prompting mechanism. Using the Basic prompt (namely, just providing a questionnaire with no further instructions), the LLM was likely to either generate negligible variance across generated personas, or generate internally inconsistent outputs (respond differently to questions about the same value). These results suggest that LLMs cannot be treated as ’individuals’ holding a coherent set of value priorities.

Prompts that endow the LLM with a “personality” improved the consistency of each specific value profile, to varying degrees. The value hierarchy was consistently found across prompts, indicating that at the mean level, LLMs can simulate value rankings of human populations. More variability between conditions was found in the internal consistency metric, with the Values Anchor prompt providing the best consistency, and the Names prompt a low consistency that would be considered unacceptable in a human sample. Finally, the Value Anchor prompt also best modeled the inter-value correlations. This is arguably the most important metric since it allows analyzing consistency across values, within a single session. These results suggest that LLMs, applying suitable prompts, can produce a ’population’ of individuals, each reporting a different, but coherent set of value priorities.

One fascinating question is where the LLM learns to produce such clear profiles of values. These profiles may be implicitly learned during pre-training. Indeed, past studies indicated that values can be identified in texts, such as newspaper articles and social media. However, these values did not necessary follow the theoretical value inter-relations identified here [43, 44, 45]. Individuals who value competing values may experience stress and indecision when faced with a dilemma, resulting in gradual change in values toward a more coherent form [46, 42]. In contrast, text may very well present both sides of a dilemma and thus retain inconsistencies. Such inconsistencies were not identified because past studies relied mostly on lexical approaches [43, 44, 45]. LLMs, taking context into account, may be more likely to identify value inter-relations correctly. LLMs may also have learned to produce value profiles in the process of fine-tuning or RLHF [47]. Future research can try to distinguish these two sources of learning using careful analysis of training sources as well as evaluation of different checkpoints in the training process.

Past research into human personas sought ways to estimate the ability of the LLM to maintain a consistent persona across a conversation [34]. We establish that the unique qualities of human values, and the ample empirical knowledge collected about them, allow their use as such a method [3, 4, 48]. We suggest that known behavioral correlations in humans can be applied to assess the consistency of LLM personas. Here we focused on evaluations via a questionnaire, but one could envision more elaborate evaluations that rely on other features of human personalities.

The procedures and data produced here may have important contributions for psychological research. Investigators interested in human behavior can apply these procedures to produce datasets that simulate human samples. Future research can investigate their possible use to replicate known findings (e.g., age differences in values) or pretest novel hypotheses (e.g. associations between values and specific behaviors).

The question of values in LLMs is of course of philosophical and societal importance. Our results show that on average, these values largely reproduce international value rankings. However, small variations in value importance may have implications at the societal and individual level (e.g. gender roles [49]; entrepreneurship [50]; prosocial behavior [51]; and antisocial behavior [52]). Future work should consider the influence of these values on LLM responses, as well as on the individuals interacting with them.

The current study focused on a limited number of contexts (e.g., five prompts, two temperatures, and two models). Importantly, we found commonalities across the various contexts, beside the differences. Future studies can use these results to understand what other contexts should be investigated, to possibly further enhance the quality of the output. Another limitation is the restriction to one questionnaire (PVQ-RR) corresponding a specific value system. It will be interesting to explore other forms of probing values.