Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization

Researchers have proposed two new frameworks to improve role-playing agents and multimodal large language models. The frameworks, Psy-CoT and HyLaR, aim to enhance the complex problem-solving capabilities of these models.

The Psy-CoT framework, proposed in a paper submitted on 25 Jun 2026^[1], uses a psychology-grounded chain-of-thought approach to improve role-playing agents. It decomposes pre-response reasoning into three role-specific steps: Interaction Perception, Psychological Empathy, and Logical Construction. This allows the model to think dynamically from a given profile rather than just mimicking surface patterns. The paper also introduces Role-Aware Policy Optimization (RAPO), which uses profile-token mutual information to weight gradients asymmetrically. Experiments on CoSER, CharacterBench, and CharacterEval showed that Psy-CoT outperformed existing role-playing CoT methods. Another paper, submitted on 22 Apr 2026^[2], discussed HyLaR, a latent reasoning paradigm that internalizes visual states to overcome limitations in adapting Chain-of-Thought (CoT) to vision. HyLaR was found to outperform standard MLLMs and state-of-the-art latent reasoning approaches across various benchmarks. Both papers were submitted in 2026^[1][2].