Reasoning Controllability of Large Language Models: Tension Between Compliance and Sensibility

This article focuses on the controllability issue in the reasoning process of Large Language Models (LLMs), with a core discussion on the tension between compliance (strictly following instructions) and sensibility (logical consistency and factual accuracy). It aims to provide a theoretical framework and practical insights for improving the reasoning controllability of models.

Original Author: Xingwei Tan Source Platform: GitHub Publication Time: May 26, 2026 Original Link: https://github.com/Xingwei-Tan/compliance_sensibility

Large Language Models (LLMs) have made breakthrough progress in the field of natural language processing, but their controllability in reasoning remains a core concern for academia and industry. As model capabilities continue to enhance, ensuring that models maintain logical sensibility while following instructions has become an urgent challenge to address.

Traditional LLM training methods often focus on improving the model's instruction-following ability, i.e., making the model "comply" with user input. However, such pure compliance may lead the model to mechanically execute even when faced with unreasonable or contradictory instructions, lacking judgment on the rationality of the instructions. This raises a key question: How to find a balance between compliance and sensibility?

This study proposes two core concepts, "Compliance" and "Sensibility", to describe the two different dimensions of reasoning controllability in LLMs.

Compliance Dimension

Compliance refers to the degree to which the model follows user instructions. A model with high compliance will strictly generate outputs according to user prompts, even without deviating when the instructions have obvious problems. This feature is valuable in certain scenarios, such as when precise execution of specific tasks is needed. However, excessive compliance may cause the model to lack critical thinking and fail to identify logical flaws or potential risks in the instructions.

Sensibility Dimension

Sensibility focuses on the logical consistency and factual accuracy of the model's output. A model with high sensibility can make independent judgments when generating content and refuse to execute obviously unreasonable or harmful instructions. This ability is crucial for ensuring the safety and reliability of model outputs, but it may also lead the model to be too "stubborn" and refuse to cooperate in situations that require flexible handling.

Achieving reasoning controllability in LLMs faces multiple technical challenges. First, there is an inherent tension between these two dimensions: over-emphasizing compliance may impair sensibility, and vice versa. Second, different application scenarios have varying demands for these two dimensions, requiring the model to have the ability to adjust dynamically.

In addition, how to quantitatively evaluate the reasoning controllability of models is also a difficult problem. Existing evaluation metrics often only capture the performance of a single dimension, making it difficult to fully reflect the model's comprehensive performance in complex scenarios.

Improving the reasoning controllability of LLMs has broad practical significance. In fields with high requirements for accuracy and safety, such as intelligent assistants, educational tutoring, and medical consultation, models need to maintain professional judgment while following user intentions. The framework proposed in this study provides a theoretical basis for understanding and improving model behavior in these applications.

For developers, understanding the tension between compliance and sensibility helps design more robust prompt strategies and model fine-tuning schemes. By consciously adjusting the model's performance in these two dimensions, the behavior patterns of the model can be optimized for different application scenarios.

This study reminds us that the capability evaluation of LLMs should not only focus on task completion rate or instruction-following degree but also consider the rationality and safety of model outputs. Future research may explore how to better integrate these two dimensions into model architecture and training objectives, developing AI systems that are both obedient and "smart".

For practitioners using LLMs, this framework provides a useful thinking tool: when designing and evaluating AI applications, it is necessary to consider both the model's response to instructions and its judgment ability on content, finding the optimal balance suitable for specific scenarios.