How Personality Traits Influence Negotiation Outcomes?
A Simulation based on Large Language Models

We consider a classical bargaining scenario in which a buyer and a seller negotiate over the price of an item or product. Typically, the buyer aims to reduce the purchase price while the seller seeks to maximize it, resulting in the competitive nature of the negotiation scenario. This is also an example of a zero-sum game in which one party’s gain leads to the other party’s loss, showing the competitive nature of the task Gibbons (1992). In our LLM-based negotiation simulations, the seller and the buyer are LLM agents. Since our goal is to study the effect of personality traits on negotiation, we define the seller and buyer agents in terms of their psychological and economic profiles as in (Eq. 1).

The psychological profiles $\psi_{s}$ and $\psi_{b}$ will be instantiated with predefined personality traits following the Big-Five model of personality. Such model decomposes human personality into five dimensions: openness (OPE), conscientiousness (CON), extraversion (EXT), agreeableness (AGR), and neuroticism (NEU) Costa Jr and McCrae (1995); John et al. (1999). Each dimension is a spectrum with negative and positive polarities. The five dimensions encompass a comprehensive range of human personality patterns. The economic profiles of the agents are reflected in their utility functions, denoted by $u_{s}$ for sellers and $u_{b}$ for buyers. A utility function is a mathematical way to describe the preferences or objectives of the agents depending on whether they are minimizing (buyer) or maximizing the price (seller) Gibbons (1992).

The seller and the buyer negotiate in a dialogue $D$ around a product. The dialogue is a sequence of $T$ utterances $D=\{d_{1},d_{2},\ldots,d_{T}\}$. Each utterance $d_{t}$ is associated with a negotiation state $s_{t}$ and the current offer price $p_{t}$.

We configure an LLM agent with specific personality traits by introducing a personality instruction (Section 3.2.1) and a negotiation objective instruction (Section 3.2.2), with in-context learning.

Using the methods in Section 3.2, we configure the buyer LLM agent and the seller LLM agent and conduct a negotiation simulation between them. The seller and buyer agents exchange offers, with the seller kick-starting the conversation with the fixed utterance “Hi, how can I help you?”. After an utterance $d_{t}$ is generated, the response is fed to the other agent as a prompt. The process continues until the termination condition is met. In this fashion, we collect a negotiation dialogue $D=\{d_{1},d_{2},\ldots,d_{T}\}$. Following Fu et al. (2023), we introduce a dialogue state detector to extract negotiation-related information from each utterance. First, we detect the negotiation state $s_{t}$ of $d_{t}$, which is one of the following states:

• •

  Offer: the agent makes a price offer.

• •

  Ponder: the agent considers whether to accept or reject the other agent’s offer.

• •

  Accept: the agent accepts the current offer.

• •

  Deal-break: the agent refuses the last offer or walks away from the negotiation.

• •

  Chit-chat: Utterances with intent not directly related to the negotiation, such as greetings.

In addition, we extract the current offer price and the strategy of $\sigma_{t}$ each utterance $d_{t}$. After generating each utterance, the dialogue state detector takes $d_{t}$ and its context (previous $h$ utterances) as input and extracts the negotiation state $s_{t}$, current offer price $p_{t}$ (if any), and strategy $\sigma_{t}$ of the current speaker in free text form. In this work, we use another LLM as the dialogue state detector.

For the termination condition of the negotiation is based on the detected negotiation states. We terminate the negotiation dialogue if an Accept or a Deal-break is reached. Also, we set a length limit of $T=T_{max}$ of the dialogue. If the length of the generated dialogue reaches this limit, we terminate the process and automatically regard it as a failed negotiation.

We mainly evaluate the negotiations in terms of utility and whether the negotiations are successful or not Baarslag et al. (2016); Lin et al. (2023). Recall that utility functions serve as mathematical tools for quantifying the quality of decision outcomes Simon (1990). In our study we adopt a number of economic metric commonly used to evaluate negotiations Baarslag et al. (2016).

First, we analyze the relationships between the negotiation outcomes and the personality traits of agents. Table 2 illustrates the Spearman’s rank correlation coefficients between the economic metrics introduced in section 4.2 and Big-Five personality traits.¹¹1The absolute value of the observed coefficients with statistical significance are all smaller than $0.3$, only indicating only a mild correlation. Along the extraversion (EXT) dimension, we observed that extroverted buyer agents tend to result in more successful negotiations ($\rho=0.072^{**}$). On the other hand, an extroverted seller agent tend to result in a slightly higher joint utility, which indicates a fairer outcome for both negotiation participants. Among the five dimensions, agreeableness (AGR) has the strongest impact on negotiation outcomes, especially for the seller. We found that being agreeable leads to a disadvantage in bargaining in terms of intrinsic utility ($\rho=-0.262^{**}$). However, agreeableness correlated positively with joint utility ($\rho=0.118^{**}$), concession rates ($\rho=0.261^{**}$), and negotiation success ($\rho=0.052^{**}$), pointing to a propensity for cooperative behavior that benefits both negotiating parties. Also, agents tend to reach an agreement with fewer number of negotiation rounds ($\rho=-0.223^{**}$) with mutually satisfying outcomes. This aligns with the positive effect of agreeableness on the negotiators’ distributive outcomes reported in Sass and Liao-Troth (2015). Aligning with previous findings Barry and Friedman (1998), we observed a positive correlation of conscientiousness (CON) and negotiation performance, especially for seller (intrinsic utility, $\rho=0.127^{**}$). Conscientiousness is also manifested in the negotiation style of not willing to concede ($\rho=-0.097^{**}$) on the seller side, and a lengthier negotiation on the buyer side ($\rho=0.083^{**}$). In psychology literature, neuroticism (NEU) and openness (OPE) are expected to play a less predominant role in negotiations Falcão et al. (2018). Similarly, we only observe some very weak correlation (with magnitude $<0.1$) along these dimensions. For instance, neurotic sellers tend to make less concessions ($\rho=-0.074^{**}$) and have a slightly higher utility gain ($\rho=0.064^{**}$), while open-minded sellers lead to enhanced joint gain ($\rho=0.063^{**}$). These results points to how personality traits could impact the outcomes of negotiations with different effects observed based on the role of the negotiator.

So far, the buyer and the seller were defined on the basis of personality traits and economic preferences. We now look at their strategies and the consequent effects on the joint negotiation outcomes. This investigation could answer questions of the type “What would be the optimal strategy for a buyer, who happens to be neurotic, when facing a disagreeable and conscientious seller?”.

At each around, each agent adopts a specific strategy $\sigma_{t}$ that results in some price $p_{t}$, inferred from state $s_{t}$. Given that the outcome of a successful negotiation is the final price $p_{deal}$, we are interested in the relationship between the adopted strategies and the joint utility of the agreement price, $u_{sb}(p_{deal})$. We manually inspected the strategies that occur more than $20$ times within the entire negotiations and construct a sets of strategy categories the agents. We then conducted a linear regression between the strategy categories and the resulting joint utility values as illustrated in Figure 2.

As shown on the right side of the Figure, collaborative strategies exhibit positive effects on the joint gain of both negotiators. For instance, accommodating ($\beta=0.125$) and conceding ($\beta=0.093$) strategies contribute positively to joint utility. Conversely, assertive and aggressive strategies ($\beta=-0.261$) negatively impact joint utility. This suggests that collaborative strategies are more beneficial for achieving mutual gains, while assertiveness diminishes joint gain. The left side of Figure 2 illustrates how these strategies depend on certain personality traits. We found that assertive strategies are more commonly expected from disagreeable agents (AGR-). On the other hand, neurotic traits (NEU+) do not lead to concessions, whereas a lack of neurotic traits (NEU-) leads to concessions. Finally, agents with neurotic traits seem to employ more empathetic strategies, while by contrast, agents with fewer neurotic traits will refrain from them.

In addition to the joint case illustrated in Figure 2, we have also looked at the individual gains of the buyer and the seller using their intrinsic utility functions $u_{b}(p_{deal})$ and $u_{s}(p_{deal})$. Again, we start by constructing the strategy sets of the agents $k\in\{s,b\}$, namely $\Sigma_{s}$ and $\Sigma_{b}$. We found that $\Sigma_{b}\subseteq\Sigma_{s}$, indicating that sellers have additional strategies such as attempting to emphasize the target price. Generally, at time $t$, agent $k$ adopts a strategy $\sigma_{k,t}\in\Sigma_{k}$, resulting in some price $p_{t}$, inferred from state $s_{t}$. Given that the outcome of a successful negotiation is the final price $p_{deal}$, we are interested in the relationship between the strategies of the agents $\Sigma_{s}$ and $\Sigma_{b}$ and the utility values of the deal price of the buyer $u_{b}(p_{deal})$ and seller $u_{s}(p_{deal})$. To this end, we conducted a regression between strategies and the the quantities, illustrated in Figure 3, aggregated across products and personality traits. Labels on the $y$ axes are the predominant, unified strategies of players.

Figure 3(a) (right) illustrates a nuanced impact of strategies on the seller’s utility gain, highlighting a spectrum of strategies from cooperative to non-cooperative tactics (soft to hardball). Strategies positioned at the assertive end of the spectrum, such as assertiveness ($0.23$) exhibit a positive effect leading to an enhanced utility gain $u_{s}(p_{deal})$. On the cooperative end, strategies such as concession ($-0.11$) or empathy ($-0.13$) show negative impacts on the seller’s utility. These results suggest that assertive strategies are more beneficial for sellers, while cooperative strategies tend to diminish their utility gain. We now propose to look at the relatedness of the strategies of the agents to their personality traits using a Pearson’s Chi-square test. The heat-map cells on Figure 3(a) (left) illustrate the positive (red) and negative (blue) strength of dependence between the traits (x axis) and strategies (y axis). We could see for example, that disagreeableness (AGR-) aligns with non-collaborative strategies, while neuroticism (NEU) aligns with aggressiveness.

For the buyer (Figure 3(b) (right)), collaborative strategies such as softness ($0.06$) and empathy ($0.03$) exhibit positive effects on deal utility $u_{b}(p_{deal})$, indicating that strategies substantially enhance the buyer’s gains. Concession shows moderate negative impacts ($-0.01$). Assertiveness ($-0.11$) is associated with the most significant reduction in the utility. Collaborative strategies seem to also be more beneficial for buyers, while assertiveness tend to diminish their utility gain. Figure 3(b) (left), shows that a neurotic buyer (NEU+) tends to concede less and use empathy, while a less neurotic buyer (NEU-) will concede more and use less emphatic strategies.

From Figures 3(b) and 3(a), it is possible to recover the strategic asymmetry that is characteristic of our bargaining game. Some strategies have opposing effects on the intrinsic utilities of the agents. For instance, assertiveness has a coefficient of $-0.1$ for the buyer and $0.23$ for the seller. Assertiveness, however, is detrimental to the joint utility as illustrated by the coefficient $-0.45$ on Figure 2.

We conducted qualitative analysis of the negotiation simulation results. Table 3 shows two distinct examples of dialogues generated in the negotiation simulations.

In the example on the first row, the seller and the buyer are negotiating over a stereo speaker. When the buyer asked about the price, the seller provides a price ($\$80$) that is higher than the listing price ($\$50$), and ends up striking a deal at $\$60$, which is higher than the listing price. We observed this kind of deceiving behavior in the negotiations involving disagreeable agents. The second example is a negotiation on a second-hand iPhone. We can see that buyer tries to emotionally appeal to the seller by mentioning her ‘my elderly parent’ but the seller is not persuaded and adopts a ‘take-it-or-leave-it’ strategy. Despite the constant effort of the buyer to find a common ground with the seller, the later is not willing to change the offer price ($\$160$). We observe more similar ‘hardball’ negotiation strategies with disagreeable and close-minded agents. At the end of the negotiation, the buyer concedes and accept the seller’s offer. The above examples showcase a range of negotiation behaviors such as deception, emotional appeal, hard-headed behavior Baarslag et al. (2016). This illustrates how specific personality traits influence negotiation styles.