LLMs Are Too Smart to Be Average: Controlling LLM Proficiency via Guided Decoding

Dongliang Guo, Weili Shi, Sheng Li

Keywords: simulation, guided decoding

Abstract: Large Language Models (LLMs) are increasingly being used to simulate human behavior in applications such as educational technology, user modeling, and human-AI interaction. However, LLMs often default to expert-level reasoning, even when prompted to simulate individuals with limited or average proficiency. This misalignment limits their ability to realistically simulate users with diverse proficiency. In this work, we propose \textbf{Guided Decoding}, a decoding-time method for controlling the reasoning proficiency of LLMs during inference. Our approach fuses token-level logits from two sources: a reference prompt that elicits expert-level reasoning, and one or more guidance prompts that induce suboptimal reasoning patterns. By adjusting the contribution of these signals, our Guided Decoding enables fine-grained control over the model’s reasoning behavior. Experiments on multiple question answering benchmarks demonstrate that our method not only modulates accuracy, but also influences the reasoning process, enabling more faithful simulation of users across a spectrum of proficiency levels.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 10089