Parental Cooperation from Swiss Child Protection Reports: Reasoning Language Models Outperform Retrieval-Augmented Generation

Purpose: Assessing parental cooperation during child protection services (CPSs) interventions raises distinctive analytical challenges due to ambiguous, conflicting information. We compared reasoning language models (RLMs) to retrieval-augmented generation (RAG) in classifying parental cooperation from Swiss CPS reports. Methods: RLMs of three parameter sizes (255B, 32B, 4B) were benchmarked against a RAG-based approach using a consensus dataset of 100 reports independently classified by two expert human reviewers. Results: The largest RLM achieved the highest accuracy (89%), a nine-point improvement over RAG (80%), approaching expert inter-rater reliability (κ = 0.76). Accuracy was higher for mothers (93%) than fathers (85%), mirrored by expert reviewers. Applied to the full corpus, 31% of cases ( n = 3,900) had at least one parent with documented noncooperation. Discussion : RLMs outperformed RAG and approached expert reliability in classifying parental cooperation. Gender differences likely reflect documentation biases rather than model limitations, corroborating the stronger professional focus on mothers.