[Paper Review] Multi-Agent Causal Reasoning for Suicide Ideation Detection Through Online Conversations

Jun Li, Xiangmeng Wang+7 more|arXiv (Cornell University)|Feb 27, 2026

Mental Health via Writing0 citations

TL;DR

The paper introduces MACR, a two-agent framework that uses counterfactual reasoning and front-door causal adjustment to predict suicide risk from online conversation trees, outperforming baselines.

ABSTRACT

Suicide remains a pressing global public health concern. While social media platforms offer opportunities for early risk detection through online conversation trees, existing approaches face two major limitations: (1) They rely on predefined rules (e.g., quotes or relies) to log conversations that capture only a narrow spectrum of user interactions, and (2) They overlook hidden influences such as user conformity and suicide copycat behavior, which can significantly affect suicidal expression and propagation in online communities. To address these limitations, we propose a Multi-Agent Causal Reasoning (MACR) framework that collaboratively employs a Reasoning Agent to scale user interactions and a Bias-aware Decision-Making Agent to mitigate harmful biases arising from hidden influences. The Reasoning Agent integrates cognitive appraisal theory to generate counterfactual user reactions to posts, thereby scaling user interactions. It analyses these reactions through structured dimensions, i.e., cognitive, emotional, and behavioral patterns, with a dedicated sub-agent responsible for each dimension. The Bias-aware Decision-Making Agent mitigates hidden biases through a front-door adjustment strategy, leveraging the counterfactual user reactions produced by the Reasoning Agent. Through the collaboration of reasoning and bias-aware decision making, the proposed MACR framework not only alleviates hidden biases, but also enriches contextual information of user interactions with counterfactual knowledge. Extensive experiments on real-world conversational datasets demonstrate the effectiveness and robustness of MACR in identifying suicide risk.

Motivation & Objective

• Motivate robust suicide risk detection in online conversations beyond logged interactions.
• Mitigate unobserved confounders such as user conformity and copycat behavior.
• Develop a scalable framework that enriches context with counterfactual reasoning.
• Leverage front-door adjustment to provide unbiased causal estimates.
• Demonstrate effectiveness on real-world conversational datasets.

Proposed method

• Propose a two-agent MACR framework: Reasoning Agent generates counterfactual mediator M using four sub-agents based on Paul-Elder critical thinking and cognitive appraisal theory.
• Implement a Bias-aware Decision-Making Agent that performs front-door adjustment to estimate P(Y|do(X)) via P(m|X) and P(Y|do(m)).
• Decompose front-door adjustment into two computational parts and realize them with specialized agents.
• Use clustering (E5-large-v2 encoder + K-means) to consolidate counterfactual inferences into representative mediator nodes.
• Approximate P(Y|do(m)) with a sampling-based, stratified in-context prompting strategy to mitigate distributional bias.
• Compare against diverse baselines including traditional risk models, debiased LMs, LLMs, and GNNs on two real-world datasets.

Experimental results

Research questions

• RQ1 Can unobserved confounders in online suicidal discourse be addressed via front-door adjustment with a mediating variable?
• RQ2 Does enriching context with counterfactual psychological mediators improve suicide risk prediction over logged interactions alone?
• RQ3 How does multi-agent collaboration (reasoning plus bias-aware decision making) affect predictive performance and robustness?
• RQ4 Is the proposed front-door approach superior to standard back-door debiasing strategies in this domain?

Key findings

• MACR achieves state-of-the-art Weighted-F1 on both datasets (PFA: 0.3768; SCT: 0.5108).
• On the SCT dataset, MACR outperforms the second-best model by 12.3% in Weighted-F1.
• Debiased language models improve but do not reach MACR’s performance, highlighting limitations of back-door debiasing alone.
• Ablation shows the Reasoning Agent contributes the most to performance, with multi-agent collaboration and front-door adjustment both essential.
• Front-door-based prompting outperforms In-Context Learning and Chain-of-Thought prompting for this task.