Interpolative Decoding: Exploring the Spectrum of Personality Traits in LLMs

TL;DR

The paper tackles the overhead and replicability challenges of emulating human personality in LLMs by introducing interpolative decoding, a parametric method that interpolates between extremal personality prompts using a continuous lambda. It demonstrates that both Big Five and HEXACO traits can be modulated along continuous spectra, and that such modulation shapes decision-making in economic games and information integration tasks, aligning with human psychological findings. The authors also introduce twinning, an optimization-based approach to mirror individual human behavior by tuning lambda across scenarios, and provide preliminary evidence that this can approximate single-subject decisions in collaborative games. While promising, the work notes limitations in dimensionality, the depth of decoding, and generalizability to non-human agents, outlining a path toward broader, higher-fidelity, and ethically aware applications in cognitive and social science research.

Abstract

Recent research has explored using very large language models (LLMs) as proxies for humans in tasks such as simulation, surveys, and studies. While LLMs do not possess a human psychology, they often can emulate human behaviors with sufficiently high fidelity to drive simulations to test human behavioral hypotheses, exhibiting more nuance and range than the rule-based agents often employed in behavioral economics. One key area of interest is the effect of personality on decision making, but the requirement that a prompt must be created for every tested personality profile introduces experimental overhead and degrades replicability. To address this issue, we leverage interpolative decoding, representing each dimension of personality as a pair of opposed prompts and employing an interpolation parameter to simulate behavior along the dimension. We show that interpolative decoding reliably modulates scores along each of the Big Five dimensions. We then show how interpolative decoding causes LLMs to mimic human decision-making behavior in economic games, replicating results from human psychological research. Finally, we present preliminary results of our efforts to ``twin'' individual human players in a collaborative game through systematic search for points in interpolation space that cause the system to replicate actions taken by the human subject.

Figures (5)

• Figure 1: Interpolative decoding enables a generative model to approximate behavior between two character extremes based on fixed textual descriptions.
• Figure 2: Correlations with Big Five scores between control (low lambda) and Big Five (high lambda) prompts, for contrastive (left) and mixture decoding (right).
• Figure 3: Dictator game payouts against interpolated level of HEXACO Traits (low to high).
• Figure 4: Information integration results: Probability of following suggested action decreases as $\lambda$ starts to favor tactical over social information (top left). Mentions of pro-social collaborative terms (top right) and other player (bottom left) decrease with $\lambda$ while the non-social cue tactical target (bottom right) increases.
• Figure 5: We train a regressor that given the trait extrema and the observed response, predicts the $\lambda$ that would have produced that response (right). Training data is collected by running multiple traits and $\lambda$s to the interpolative decoding LLM and recording the responses (left).