Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents
Yuxin Liu, Mingye Zhu, Siyuan Liu, Bo Hu, Lei Zhang
TL;DR
This paper introduces a theory-driven, inference-time method called Persona Dynamic Decoding (PDD) that dynamically estimates persona importance to improve role-playing agents' adherence to personas in social simulations.
Contribution
It proposes a novel, adaptive persona management framework that estimates context-dependent importance and integrates it into decoding, addressing static prompt limitations.
Findings
Improves utterance consistency in role-playing agents.
Enhances behavioral fidelity to defined personas.
Demonstrates effectiveness through extensive experiments.
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…
Peer Reviews
Decision·ICLR 2026 Poster
1. This paper addresses a practical problem — controlling persona adherence in inference-time, without fine-tuning. 2. Experimental setup includes multiple datasets (CharacterEval, Beyond Dialogue, PERSONALITYBENCH). 3. The method is described clearly, and the PIE and PIA stages are straightforward.
1. For the Persona Importance Estimation (PIE) module in assumption (Proposition 3.2), the authors "propose using G as an approximation of GT," where G is the model's own generated response. What if the model's generation $G$ is low quality or misaligned with the persona? Then, it seems that PIE will calculate its "importance scores" based on this output, and the model's own errors are used to guide its generation, potentially reinforcing its existing biases rather than correcting them. 2. The
- Introduces a theory-grounded, inference-time persona adaptation mechanism inspired by psychological models (CAPS), offering a fresh perspective on dynamic persona control. - The paper writes with great clarity, paper structure and visuals explainer - Strong empirical validation with both quantitative metrics and human/LLM-based evaluations demonstrating robustness and generalizability of the method.
- Dependence on subjective LLM-based judges for evaluation may introduce bias and reduce reliability. - The paper’s practical implications for broader applications beyond controlled experiments (e.g., open-domain dialogue) are not fully explored, which limits significance.
1. The paper integrates interdisciplinary theories (Cognitive-Affective Personality Systems from psychology and conditional mutual information from information theory) to underpin the PDD framework, rationally justifying dynamic persona adaptation and avoiding blind technical design. 2. PDD enables fine-tuning-free dynamic persona following at inference time: its PIE module adjusts persona attribute priority based on scenarios, while PIA modulates generation probabilities, solving existing metho
1. The robustness analysis of PIE’s core assumption (using model-generated G to approximate ground-truth GT) is insufficient; the paper fails to discuss the risk of misjudging importance scores when G deviates from persona settings, nor does it test extreme scenarios like conflicting persona attributes. 2. Fig.4 shows that increasing persona attributes leads to a win rate decline. Does this indicate that PDD cannot model scenarios with complex and multiple attributes?
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsPersona Design and Applications · Social Robot Interaction and HRI · Multimodal Machine Learning Applications