Benchmarking Implicit Suicidal Ideation in Open-Weight Large Language Models Under the Columbia-Suicide Severity Rating Scale

Abstract Conversational artificial intelligence platforms and automated psychological screening agents are increasingly deployed to alleviate catastrophic shortages of mental healthcare professionals in low-to-middle-income countries (LMICs). However, prevailing safety mechanisms and clinical natural language processing (NLP) classifiers remain almost exclusively calibrated on explicit, overt disclosures of crisis, creating a severe diagnostic blindspot when confronted with subtle, veiled, or metaphorical suicidal ideation. In this paper, we establish a comprehensive empirical benchmark evaluating open-weight Large Language Models (LLMs)—specifically Meta Llama-3-8B-Instruct and Qwen-2.5-7B-Instruct—against domain-specific discriminative encoders (Mental-RoBERTa) and commercial frontier anchors (GPT-4o) across the standardized 6-stage clinical continuum of the Columbia-Suicide Severity Rating Scale (C-SSRS). Utilizing the psychologically simulated DeepSuiMind corpus paired with the UMD Reddit Suicidality overt control distribution, we uncover a devastating "implicit gap": under standard zero-shot prompting, open-weight foundation models suffer an average 61.5% relative collapse in Macro-F1 and exhibit an alarming 22.9% High-Risk False Negative Rate, routinely misclassifying imminent suicide preparations as benign stress or passive sorrow. To resolve this life-threatening deficiency on consumer-grade hardware, we introduce a structured Clinical Few-Shot Chain-of-Thought (CoT) prompting framework and a 4-bit Parameter-Efficient Fine-Tuning (QLoRA) pipeline. Our empirical profiling across 5 random seeds demonstrates that Clinical Few-Shot CoT completely eliminates high-risk triage omissions (0.0% High-Risk FNR), restoring Macro-F1 to 0.899 (±0.053), Quadratic Weighted Kappa to 0.985 (±0.008), and stage Mean Absolute Error to 0.09, matching or surpassing commercial proprietary systems. All code, prompts, trial logs, and benchmark datasets are released as an open-source clinical auditing suite.