Sparks of Rationality: Do Reasoning LLMs Align with Human Judgment and Choice?

TL;DR

This work probes whether reasoning-enabled LLMs align with human judgment and decision biases by jointly benchmarking compliance with four rational-choice axioms and neuronal-like affective distortions. It introduces two emotion-steering methods—in-context priming and representation-level steering—to elicit emotion-driven changes, and evaluates across risk, belief formation, norm enforcement, and prosocial decision domains. The findings show that explicit thinking reliably enhances rationality and EV-maximizing tendencies, while affective steering can produce directionally human-like shifts that vary in magnitude and reliability depending on the steering method and model family. A key takeaway is a tension between accurate human-like simulation and robust, unbiased decision support: ICP yields strong, sometimes exaggerated shifts toward human-like behavior, whereas RLS offers graded, more controllable patterns with clear safety implications. The proposed benchmarking framework supports auditing and shaping LLM decision behavior in real deployments, with practical relevance for medicine, hiring, finance, and policy where steering and reasoning interact with safety and fairness constraints.

Abstract

Large Language Models (LLMs) are increasingly positioned as decision engines for hiring, healthcare, and economic judgment, yet real-world human judgment reflects a balance between rational deliberation and emotion-driven bias. If LLMs are to participate in high-stakes decisions or serve as models of human behavior, it is critical to assess whether they exhibit analogous patterns of (ir)rationalities and biases. To this end, we evaluate multiple LLM families on (i) benchmarks testing core axioms of rational choice and (ii) classic decision domains from behavioral economics and social norms where emotions are known to shape judgment and choice. Across settings, we show that deliberate "thinking" reliably improves rationality and pushes models toward expected-value maximization. To probe human-like affective distortions and their interaction with reasoning, we use two emotion-steering methods: in-context priming (ICP) and representation-level steering (RLS). ICP induces strong directional shifts that are often extreme and difficult to calibrate, whereas RLS produces more psychologically plausible patterns but with lower reliability. Our results suggest that the same mechanisms that improve rationality also amplify sensitivity to affective interventions, and that different steering methods trade off controllability against human-aligned behavior. Overall, this points to a tension between reasoning and affective steering, with implications for both human simulation and the safe deployment of LLM-based decision systems.

Figures (70)

• Figure 1: We study LLM decision-making, focusing on reasoning models under in-context emotion manipulation or via vector injection. We differentiate between (Top) neutral thinking, (Middle)thinking about emotions and (Bottom)emotional thinking.
• Figure 2: Rationality compliance for (Top) different LLMs and (Bottom) emotional Qwen3 with different steering strengths. Scores indicate the proportion of decision instances that satisfy axioms of rationality, providing a summary of preference coherence.
• Figure 3: Decision Domains and the effects of induced emotions reported in human studies.
• Figure 4: Risk preference curves: fitted choice probability versus EV difference, comparing neutral and emotion ICP/RLS steering.
• Figure 5: Fitted Prelec probability-weighting functions, comparing neutral and fear steering.