Emotion-Steering Methods Affect Rationality in Large Language Models, Study Finds

Global: Emotion-Steering Methods Affect Rationality in Large Language Models, Study Finds

On January 2026, a team of researchers posted a new preprint on arXiv that examines whether large language models (LLMs) display patterns of rationality and bias comparable to human decision‑making. The study evaluates multiple LLM families across benchmarks that test core axioms of rational choice as well as classic behavioral‑economic scenarios where emotions typically shape judgments. Its purpose is to inform the safe deployment of LLMs in high‑stakes contexts such as hiring, healthcare, and economic forecasting.

Assessing Core Rationality

The authors first measured how well LLMs adhere to established rational‑choice principles, including consistency, transitivity, and expected‑value maximization. Results indicate that, when prompted to engage in deliberate “thinking,” models tend to produce choices that align more closely with these axioms, suggesting that prompting strategies can enhance logical reasoning.

Emotion‑Steering Techniques

To probe affect‑driven distortions, the researchers applied two distinct steering methods: in‑context priming (ICP), which inserts emotionally charged cues into the prompt, and representation‑level steering (RLS), which adjusts internal model representations toward specific affective states. Both techniques aim to simulate how emotions might bias human judgments.

Impact of Deliberative Thinking

Across all experimental conditions, the study finds that encouraging models to “think step‑by‑step” consistently improves rational outcomes. However, the same prompting also amplifies the models’ sensitivity to emotional cues, leading to larger deviations from rational benchmarks when steering is applied.

Comparing ICP and RLS

In‑context priming produced pronounced, often extreme shifts in model behavior, making the direction of change easy to predict but difficult to calibrate for nuanced applications. By contrast, representation‑level steering generated more psychologically plausible patterns that resembled human affective bias, yet the effects were less reliable and varied across model families.

Implications for Human Simulation

The findings suggest a trade‑off between controllability and human‑aligned behavior: methods that offer precise steering may generate unrealistic emotional responses, while those that mimic human affect tend to be less predictable. This tension has direct relevance for projects that aim to model human decision‑making or to embed LLMs in environments where emotional intelligence is valued.

Considerations for Safe Deployment

According to the authors, the dual influence of reasoning prompts and affective steering highlights the need for careful design of LLM‑based decision systems. Ensuring that models remain rational while avoiding unintended emotional bias will be crucial for applications that impact real‑world outcomes.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.