Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents

TL;DR

This work introduces the Persona Dynamic Decoding (PDD) framework, a novel, theory-driven method that dynamically estimates context-dependent persona importance and integrates it into weighted reward-guided decoding, enabling inference-time persona following.

Abstract

The utility of Role-Playing Language Agents in sociological research is growing alongside the adoption of Large Language Models. For realism in social simulation, these agents must adhere to their personas defined by character profiles, yet existing strategies-static prompt engineering or costly fine-tuning-fail to adapt personas to dynamic scenarios. Psychological theories, such as the Cognitive-Affective Personality Systems, provide a crucial explanation for this failure: a persona's influence on behavior is not static but varies with the scenarios. This context-dependence highlights the critical need for adaptive persona management. To address this gap, we propose a novel, theory-driven method that dynamically estimates context-dependent persona importance and integrates it into weighted reward-guided decoding, enabling inference-time persona following. Specifically, we introduce the Persona Dynamic Decoding (PDD) framework, which consists of two key components: (1) Persona Importance Estimation (PIE) module, which dynamically quantifies the contextual importance of persona attributes without requiring ground-truth supervision; and (2) Persona-Guided Inference-Time Alignment (PIA) paradigm, which leverages these importance scores to construct weighted multi-objective rewards and modulate generation probabilities during inference. Extensive experiments show the effectiveness of our method in utterance consistency and behavioral fidelity.

Figures (8)

• Figure 1: Overview of the proposed PDD. Given a full prompt $T$ including character profile $P$, scenario context $C$ and a query $x$, PDD first dynamically estimates persona importance $I_i$ for each persona attribute $w_i$. Then we compute persona-guided rewards $r_i(T,y_{<t})$ by measuring the divergence between persona-constrained and unconstrained probability distributions. These individual rewards are then integrated into a normalized reward signal $R_{\text{norm}}$, which emphasizes more important attributes through adaptive weighting. Finally, the aligned policy $p_r$ can be derived through reward maximization, thereby generating persona following responses.
• Figure 2: Overall assessment of PIE for general character task. Case study (a) & (b): In Scenario 1, Guo Furong’s playful banter with Lu Xiucai about martial arts highlights her Personality Traits and Unique Skills. In Scenario 2, her guidance to Tong Xiangyu foregrounds Life View and Educational Views, illustrating context-dependent persona relevance. Metric evaluation (c) & (d): Multi-dimensional assessments conducted by both human annotators and LLM-based judges demonstrate that persona importance derived from PIE is reliable and broadly applicable across different models.
• Figure 3: Direct comparison with the baselines for specific personality task on PERSONALITYBENCH.
• Figure 4: Ablation study of the effect of the number of persona attributes on CharacterEval.
• Figure 5: Overall assessment of PIE for specific personality task. Case Study (a) & (b): In Scenario 1, the agent faces a moral dilemma with top keywords like empathy and great friend reflecting the emotional context. Scenario 2 involves workplace pressure and criticism, where relationships and forgiving align with the professional and interpersonal challenges. Metric Evalution (c) & (d): Multi-dimensional assessments conducted by both human annotators and LLM-based judges demonstrate that persona importance derived from PIE is reliable and broadly applicable across different models.

Theorems & Definitions (6)

• Proposition 3.1
• Remark 1
• Proposition 3.2
• Remark 2
• Proposition 3.3
• Remark 3