How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities
This work addresses the problem of ensuring safe and controllable behavior in Large Language Models for socially sensitive domains, which is crucial for their reliable deployment.
The authors tackled the problem of unpredictable behaviors in Large Language Models, finding that control often degrades at finer-grained levels. Their SteerEval benchmark provides a framework for evaluating LLM controllability across three domains.
Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerEval, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.