OpenNovelty.org

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

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.5Download Report PDF
Agent-Based Simulation; Role-Playing Language Agents; Persona Following; Inference-Time Alignment
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 Persona Dynamic Decoding (PDD) framework that 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.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces a Persona Dynamic Decoding (PDD) framework that dynamically adjusts persona attribute importance during inference, grounded in Cognitive-Affective Personality Systems theory. It resides in the Dynamic Persona Adaptation leaf, which contains only three papers total, indicating a relatively sparse research direction within the broader taxonomy of fifty papers. This leaf focuses specifically on methods that adjust persona influence contextually during generation, distinguishing it from static prompt engineering or offline training approaches that dominate neighboring branches.

The taxonomy reveals that Dynamic Persona Adaptation sits within Dialogue Generation and Behavioral Alignment, adjacent to leaves addressing emotion-aware role-playing, character consistency mechanisms, and retrieval-augmented approaches. While neighboring branches like Profile-Based Persona Representation (five papers) and Personality-Driven Persona Construction (four papers) focus on static persona encoding, and Training and Alignment Paradigms addresses offline optimization, this leaf uniquely targets inference-time adaptation. The scope note explicitly excludes static methods, positioning the work at the intersection of real-time behavioral adjustment and persona fidelity maintenance.

Among eighteen candidates examined, none clearly refute the three core contributions: the PDD framework (two candidates examined), the PIE module for unsupervised importance estimation (ten candidates), and the PIA inference-time alignment paradigm (six candidates). The PIE module received the most scrutiny, yet no overlapping prior work emerged from this limited search. This suggests that within the top semantic matches and their citations, the specific combination of dynamic importance quantification without ground-truth supervision and weighted reward-guided decoding appears relatively unexplored, though the modest search scope leaves open the possibility of relevant work beyond these eighteen papers.

Given the sparse population of the Dynamic Persona Adaptation leaf and the absence of refuting candidates among eighteen examined papers, the work appears to occupy a less-crowded niche within role-playing agent research. However, the limited search scale—eighteen candidates from semantic retrieval—means this assessment reflects only a narrow slice of the literature. The taxonomy structure indicates growing interest in dynamic adaptation mechanisms, yet the specific theory-driven approach to context-dependent persona weighting may represent a novel angle within this emerging direction.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
18
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: dynamic persona following in role-playing language agents. The field centers on enabling language models to adopt and maintain consistent character identities during interaction, spanning several interconnected branches. Persona Representation and Modeling addresses how character traits, memories, and psychological profiles are encoded and retrieved, with works like Characterbox[8] and PersonaAgent GraphRAG[29] exploring structured knowledge bases and graph-based retrieval. Dialogue Generation and Behavioral Alignment focuses on producing utterances that reflect these personas authentically, including dynamic adaptation mechanisms that adjust character behavior in response to conversational context. Training and Alignment Paradigms investigates methods such as reinforcement learning and profile-dialogue alignment (Profile Dialogue Alignment[4]) to improve persona consistency. Evaluation and Benchmarking provides datasets and metrics—exemplified by RoleRMBench[45] and DMT RoleBench[12]—to measure fidelity and coherence. Multimodal Role-Playing extends these ideas to voice and visual modalities (VoxRole[14]), while Application-Specific Role-Playing Systems targets domains like gaming (Dynamic NPC Dialogs[3], Personalized Quest Generation[5]) and professional simulation (Simulating Professional Workplaces[18]). Prompt Engineering and Interaction Patterns examines how carefully designed prompts and conversational scaffolds guide agents toward desired role behaviors. A particularly active line of work explores how agents can dynamically adjust persona emphasis as dialogue unfolds, balancing static character profiles with real-time contextual cues. Dynamic Importance Estimation[0] sits squarely in this space, proposing mechanisms to weigh different persona attributes adaptively rather than treating all traits uniformly. This contrasts with more static approaches like InCharacter[16] or RoleCraft GLM[13], which rely on fixed persona encodings, and complements recent efforts in dynamic context adaptation (Dynamic Context Adaptation[44]) and flexible persona frameworks (DPRF[46]). Neighboring works such as HonkaiChat[25] and Psyplay[1] also address persona consistency in interactive settings, yet Dynamic Importance Estimation[0] distinguishes itself by explicitly modeling the salience of persona elements over time. This emphasis on adaptive weighting reflects a broader trend toward more nuanced, context-sensitive role-playing agents that can navigate the trade-off between character fidelity and conversational fluidity.

Claimed Contributions

Persona Dynamic Decoding (PDD) framework

A framework that dynamically adapts persona importance to varying scenarios and guides generation without fine-tuning. It consists of two components: Persona Importance Estimation (PIE) for quantifying contextual importance of persona attributes, and Persona-Guided Inference-Time Alignment (PIA) for modulating generation probabilities during inference.

2 retrieved papers
Persona Importance Estimation (PIE) module

A module that quantifies the influence of each persona attribute by assessing Conditional Mutual Information using only inference-time log probabilities, eliminating reliance on ground-truth supervision. The authors theoretically show that model-generated outputs provide a reliable basis for deriving importance rankings.

10 retrieved papers
Persona-Guided Inference-Time Alignment (PIA) paradigm

A paradigm that uses importance scores from PIE to construct weighted multi-objective rewards and modulate token-level generation probabilities during decoding. It formulates multi-persona alignment as a normalized reward function that preserves hierarchical structure of persona attributes without requiring training.

6 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Persona Dynamic Decoding (PDD) framework

A framework that dynamically adapts persona importance to varying scenarios and guides generation without fine-tuning. It consists of two components: Persona Importance Estimation (PIE) for quantifying contextual importance of persona attributes, and Persona-Guided Inference-Time Alignment (PIA) for modulating generation probabilities during inference.

Contribution

Persona Importance Estimation (PIE) module

A module that quantifies the influence of each persona attribute by assessing Conditional Mutual Information using only inference-time log probabilities, eliminating reliance on ground-truth supervision. The authors theoretically show that model-generated outputs provide a reliable basis for deriving importance rankings.

Contribution

Persona-Guided Inference-Time Alignment (PIA) paradigm

A paradigm that uses importance scores from PIE to construct weighted multi-objective rewards and modulate token-level generation probabilities during decoding. It formulates multi-persona alignment as a normalized reward function that preserves hierarchical structure of persona attributes without requiring training.

Table of Contents