Junho Na

Daiki E. Matsunaga, Junho Na, Tri Wahyu Guntara et al.

Cooperative tasks in Multi-Agent Reinforcement Learning (MARL) require agents to collectively maximize a shared return. Under the Centralized Training with Decentralized Execution (CTDE) paradigm, policy gradients have remained difficult to compute directly. Prior methods largely follow two approaches: independent factorized updates with centralized critics, which lack general joint-improvement guarantees without value decomposition assumptions, or alternating best-response updates, which can converge to suboptimal Nash Equilibria. In this paper, we show the joint policy gradient admits an exact decentralized decomposition of per-agent terms, each formed from per-agent score functions and decentralized critics. Based on this decomposition, we develop Agent-Chained Policy Optimization (ACPO), where actors are trained independently, with their updates together constituting a single step on the joint policy gradient. Central to this result is a serialized view of the simultaneous joint decision in which agents commit actions one at a time, each conditioning on a belief over preceding actions. The belief acts as the coordination mechanism which ties the independent per-agent updates into a joint gradient step. We evaluate ACPO on Multi-Robot Warehouse, SMACv2, and MA-MuJoCo, where it outperforms strong baselines, with the gap widening as the number of agents grows.

Jae-hee So, Joonhwan Chang, Eunji Kim et al.

Background: Advancements in large language models (LLMs) have opened new possibilities in psychiatric interviews, an underexplored area where LLMs could be valuable. This study focuses on enhancing psychiatric interviews by analyzing counseling data from North Korean defectors who have experienced trauma and mental health issues. Objective: The study investigates whether LLMs can (1) identify parts of conversations that suggest psychiatric symptoms and recognize those symptoms, and (2) summarize stressors and symptoms based on interview transcripts. Methods: LLMs are tasked with (1) extracting stressors from transcripts, (2) identifying symptoms and their corresponding sections, and (3) generating interview summaries using the extracted data. The transcripts were labeled by mental health experts for training and evaluation. Results: In the zero-shot inference setting using GPT-4 Turbo, 73 out of 102 segments demonstrated a recall mid-token distance d < 20 in identifying symptom-related sections. For recognizing specific symptoms, fine-tuning outperformed zero-shot inference, achieving an accuracy, precision, recall, and F1-score of 0.82. For the generative summarization task, LLMs using symptom and stressor information scored highly on G-Eval metrics: coherence (4.66), consistency (4.73), fluency (2.16), and relevance (4.67). Retrieval-augmented generation showed no notable performance improvement. Conclusions: LLMs, with fine-tuning or appropriate prompting, demonstrated strong accuracy (over 0.8) for symptom delineation and achieved high coherence (4.6+) in summarization. This study highlights their potential to assist mental health practitioners in analyzing psychiatric interviews.