Counterfactual Debiasing to Enhance Chinese Online Counseling Personalization

Web-based Chinese online counseling platforms, particularly those empowered by large language models (LLMs), have revolutionized mental health support. However, their effectiveness is hindered by a critical bottleneck we term Response Inertia (RI) - the models' habitual reliance on specific, repetitive responses to address client issues within similar contexts, leading to a failure to provide personalized interventions tailored to users' unique circumstances and emotional needs. This phenomenon stems from data distribution bias: the scarcity of diverse user reactions within similar contexts causes systems to develop strong priors toward rigid, one-size-fits-all patterns. To address this, we propose the Counterfactual Patching for Debiasing (CPD) framework. Unlike previous methods, CPD focuses on constructing a "Dataset Patch"(RI-Patch) to actively repair distribution defects. Specifically, CPD simulates diverse users in a hypothetical space to cover long-tail personas and employs reinforcement learning as an optimization mechanism to ensure the generated counterfactual trajectories align with human preferences. By injecting this high-quality patch into the original corpus, we rebalance the training distribution to mitigate the bias. Extensive experiments demonstrate that CPD significantly enhances personalization and user satisfaction by overcoming RI, paving the way for more effective, individualized LLM-based counseling.