Can the Environment Speak for Itself? $T^{2}$-GRPO: A Turn-Trajectory Group Relative Policy Optimization for Caregiver Agents

Researchers propose T²-GRPO, a reinforcement learning framework that optimizes large language models for dementia caregiver agents by balancing immediate patient feedback with long-term care outcomes. The method uses environment-grounded rewards and safety constraints to improve emotional intelligence in AI caregiving scenarios.

T²-GRPO addresses a critical challenge in applying reinforcement learning to emotionally complex domains where traditional reward structures fail. Dementia care presents unique difficulties: patient responses are fragmented and indirect, sparse trajectory-level rewards make credit assignment problematic, and external evaluators remain expensive and unreliable. The proposed framework solves this by decoupling the learning problem into two normalized reward horizons—immediate turn-level feedback derived directly from patient state changes and longer-term trajectory evaluations. This dual-horizon approach, combined with centered-rank normalization, prevents reward collapse while preserving heterogeneous signals that reflect the nuanced nature of caregiving.

The practical implications extend beyond dementia care into any domain requiring AI agents to balance competing objectives under uncertainty. Healthcare AI development has historically struggled with safety verification and interpretability, particularly when patient autonomy and emotional wellbeing are at stake. T²-GRPO's environment-grounded reward mechanism—measuring observable changes in patient distress and resistance—creates an auditable feedback loop that reduces reliance on potentially biased external evaluators.

For the AI and healthcare technology sectors, this research demonstrates measurable progress in training agents for high-stakes human interaction scenarios. The framework's effectiveness on emotionally sensitive tasks suggests viable pathways for deploying LLM-based assistants in elder care, where labor shortages and personalization demands remain acute. The hard veto mechanism for safety constraints also reflects growing industry recognition that reward optimization alone proves insufficient for responsible AI deployment.

Developers working on caregiver systems and healthcare AI platforms should monitor refinements to this approach. Successful real-world implementation would validate whether environment-grounded rewards transfer effectively from simulators to actual patient interactions, a critical gap in current research.

• →T²-GRPO decouples reinforcement learning into immediate and long-term reward horizons to solve credit assignment in emotionally complex caregiving scenarios.
• →Environment-grounded rewards measuring observable patient state changes reduce reliance on expensive and unreliable external LLM-based evaluators.
• →Centered-rank normalization preserves heterogeneous reward signals while preventing reward collapse in multi-objective learning settings.
• →The framework includes safety constraints through binary hard veto mechanisms, addressing a critical gap in AI deployment for high-stakes human interaction.
• →Experimental results demonstrate substantial improvements over competitive baselines in dementia caregiver simulations, suggesting viability for healthcare AI applications.