How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities

Summary

This paper introduces a new way to test how well we can control the responses of large language models, which are becoming more common in everyday applications.

What's the problem?

Large language models can be unpredictable, sometimes saying things we don't want them to, or acting in ways that aren't consistent. It's hard to know *how* well we can actually steer these models to behave as intended, and current methods for controlling them don't always work well when you need very specific control over their output.

What's the solution?

The researchers created a benchmark called SteerEval. This benchmark tests control in three areas – the language used, the emotional tone (sentiment), and the model's 'personality'. It breaks down control into three levels: what the model should say, how it should say it, and specific examples of how to express it. They then used this benchmark to test existing control methods and found that control often gets weaker when you try to be more precise.

Why it matters?

This work is important because it provides a clear and organized way to evaluate and improve the safety and reliability of large language models. By understanding where current control methods fall short, researchers can develop better techniques to ensure these models behave responsibly and predictably, especially when used in sensitive situations.