TLDR: CRADLE BENCH is a new benchmark for evaluating language models’ ability to detect multi-faceted mental health crises like suicide ideation, domestic violence, and child abuse. It features 600 clinician-annotated examples with temporal labels (ongoing vs. past) and covers seven crisis types. The research evaluated 15 LLMs, showing that an ensemble approach significantly improves detection accuracy, and fine-tuned models achieve substantial performance gains. The benchmark highlights LLM weaknesses in nuanced crisis detection and provides open-source data and models for future AI safety research.
Large Language Models (LLMs) are increasingly used in sensitive areas like mental health support. However, their ability to reliably detect critical situations such as suicide ideation, domestic violence, or child abuse has been largely unexplored, and current models often fall short of human safety standards. To address this crucial gap, researchers have introduced a new benchmark called CRADLE BENCH.
CRADLE BENCH is a comprehensive tool designed to evaluate how well language models can identify various mental health crises and safety risks. Unlike previous efforts that focused on a limited number of crisis types, this benchmark covers seven distinct categories, all defined according to established clinical standards. These categories include self-harm, passive and active suicide ideation, domestic violence, rape, sexual harassment, and child abuse/endangerment. A unique feature of CRADLE BENCH is its inclusion of temporal labels, distinguishing between ongoing crises that require immediate attention and past events that provide important context.
The benchmark was meticulously constructed using real-world social media posts from crisis-related subreddits on platforms like Reddit. Mental health professionals, including licensed psychologists and social workers, meticulously annotated 600 evaluation examples and 420 development examples. This annotation process involved iterative rounds to refine guidelines and ensure high consistency, even expanding the label set to include child abuse/endangerment and differentiating suicide ideation into passive and active forms based on clinical scales.
A key aspect of CRADLE BENCH is its ability to handle multi-faceted crisis scenarios. Posts can receive multiple labels if different crisis types co-occur, reflecting the complex nature of real-world situations. For instance, a single incident might be tagged as both ‘rape’ and ‘child abuse/endangerment’ if the victim is a minor. The temporal dimension is also vital, as clinical interventions depend heavily on whether a crisis is current or historical.
The researchers evaluated 15 state-of-the-art LLMs using CRADLE BENCH. Generally, larger models performed better, and closed-source models like Gemini-2.5-Pro, Claude-4-Sonnet, and GPT-5 significantly outperformed their open-source counterparts. To further enhance performance, an innovative ensemble approach was employed, combining predictions from these three top-performing closed-source models using a majority voting scheme. This method consistently improved accuracy and was then used to automatically label a larger training dataset of around 4,000 examples, replacing costly human annotation.
Specialized crisis detection models were then fine-tuned using this ensemble-labeled data. These fine-tuned models showed notable performance gains, demonstrating that adapting LLMs to this specific task can significantly improve their ability to detect complex crisis signals in social media narratives. For example, a fine-tuned Llama 70B model achieved the best overall performance among open-source models, even surpassing some larger baselines.
However, the analysis also revealed common weaknesses in LLMs. Models often struggled with over-extending ‘child abuse’ labels to peer-to-peer conflicts rather than adult-child situations, misclassifying the severity of suicide ideation (active vs. passive), and failing to prioritize the most severe crisis type when multiple related categories were present in a single event. These insights highlight areas where future model development needs to focus on more nuanced clinical reasoning.
Also Read:
- Unpacking T2I-RiskyPrompt: A New Benchmark for Text-to-Image Model Safety
- Enhancing LLM Safety Through Smart Data Selection During Fine-Tuning
CRADLE BENCH represents a significant step forward in developing safer and more reliable AI for mental health applications. By providing a high-quality, clinician-annotated, multi-faceted, and temporally aware benchmark, along with specialized models and training data, this research lays a strong foundation for future work in this critical field. All data, models, and code are openly released to support ongoing research and development. You can find the full research paper here: CRADLE BENCH Research Paper.
