C$^3$TG: Conflict-aware, Composite, and Collaborative Controlled Text Generation

TL;DR

C3TG tackles multi-dimensional controlled text generation where attributes can conflict by coupling an LLM with lightweight attribute classifiers and a two-phase pipeline. The generation phase fuses attribute priors via a weighted KL divergence, yielding $P^{*}(x|x_{1:t-1})$ from a weighted geometric mean of priors, while the optimization phase uses an energy function $E(x)$ that combines classifier alignment $\sum_i \alpha_i |C_{A_i}(x)-T_i|$ and stability penalties $\Omega_{overlap}(x)$ to iteratively refine text through an agent-guided chain of prompts with a convergence threshold $\tau=0.025$. The framework handles 17 sub-dimensions with classifier feedback and a three-stage rewrite loop (core calibration, balancing, global fine-tuning), enabling conflict resolution and preserving fluency and diversity. Experiments on ROCStories and WritingPrompts show C3TG achieves higher attribute accuracy, reduced toxicity, and better human judgments than decoding-based and indirect CTG baselines, validating its flexible, scalable approach to multi-attribute control without extensive model modification.

Abstract

Recent advancements in large language models (LLMs) have demonstrated remarkable text generation capabilities. However, controlling specific attributes of generated text remains challenging without architectural modifications or extensive fine-tuning. Current methods typically toggle a single, basic attribute but struggle with precise multi-attribute control. In scenarios where attribute requirements conflict, existing methods lack coordination mechanisms, causing interference between desired attributes. Furthermore, these methods fail to incorporate iterative optimization processes in the controlled generation pipeline. To address these limitations, we propose Conflict-aware, Composite, and Collaborative Controlled Text Generation (C$^3$TG), a two-phase framework for fine-grained, multi-dimensional text attribute control. During generation, C$^3$TG selectively pairs the LLM with the required attribute classifiers from the 17 available dimensions and employs weighted KL-divergence to adjust token probabilities. The optimization phase then leverages an energy function combining classifier scores and penalty terms to resolve attribute conflicts through iterative feedback, enabling precise control over multiple dimensions simultaneously while preserving natural text flow. Experiments show that C$^3$TG significantly outperforms baselines across multiple metrics including attribute accuracy, linguistic fluency, and output diversity, while simultaneously reducing toxicity. These results establish C$^3$TG as an effective and flexible solution for multi-dimensional text attribute control that requires no costly model modifications.

Figures (7)

• Figure 1: Unlike existing methods that control few attributes and handle conflicts poorly, C3TG offers broad attribute control and resolves conflicts effectively.
• Figure 2: Flowchart of the overall process of the C3TG framework, including the generation phase and the optimization phase, leading to the generation of a text that meets the user's requirements.
• Figure 3: Schematic diagram of the complete Chain-of-Prompt workflow in the optimization phase of the C3TG framework.
• Figure 4: C³TG human-evaluation interface: raters choose a story, score Attribute Relevance, Fluency, and Diversity on a 1–5 scale, and export the results as a CSV file.
• Figure 5: Heat map of pair-wise Pearson correlations among our fine-grained attribute dimensions. Warm colours represent strong positive dependence, cool colours indicate strong negative dependence, and near-white cells mark dimensions that are largely independent.