Funny or Persuasive, but Not Both: Evaluating Fine-Grained Multi-Concept Control in LLMs

2026-01-26

AI Summary

Current large language models exhibit limited capability in fine-grained control over multiple stylistic or semantic concepts—such as simultaneously modulating humor and persuasiveness—and lack a systematic evaluation framework for such tasks. This work proposes the first benchmark specifically designed to assess fine-grained text control in both single- and dual-concept settings, focusing on semantically distinguishable concept pairs. By conducting a comparative analysis of prompt engineering and representation-based approaches, we formulate controllable generation tasks and introduce quantitative metrics to evaluate performance. Extensive experiments across multiple mainstream models and tasks reveal a significant performance drop when controlling dual concepts jointly, thereby exposing, for the first time, the limitations of existing prompting methods under compositional assumptions. Our framework establishes a new foundation and direction for research on multi-concept controllable text generation.

Technology Category

Natural Language Processing: GenerationMachine Learning: Large Multimodal Models (LMMs)Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

Economics, Online Markets and Human Computation: LLM based quality controls for crowd workSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Large Language Models (LLMs) offer strong generative capabilities, but many applications require explicit and \textit{fine-grained} control over specific textual concepts, such as humor, persuasiveness, or formality. Prior approaches in prompting and representation engineering can provide coarse or single-attribute control, but systematic evaluation of multi-attribute settings remains limited. We introduce an evaluation framework for fine-grained controllability for both single- and dual-concept scenarios, focusing on linguistically distinct concept pairs (e.g., persuasiveness vs.~humor). Surprisingly, across multiple LLMs and generative tasks, we find that performance often drops in the dual-concept setting, even though the chosen concepts should in principle be separable. This reveals a fundamental limitation of naive prompting-based control: models struggle with compositionality even when concepts are intuitively independent. Our framework provides systematic evidence of this gap and offers a principled approach for measuring the ability of future methods for multi-concept control.

Problem

• fine-grained control
• multi-concept control
• large language models
• compositionality
• text generation

Innovation

• fine-grained controllability
• multi-concept control
• large language models
• compositionality
• evaluation framework