Reasoning Over Recall: Evaluating the Efficacy of Generalist Architectures vs. Specialized Fine-Tunes in RAG-Based Mental Health Dialogue Systems

The deployment of Large Language Models (LLMs) in mental health counseling faces the dual challenges of hallucinations and lack of empathy. While the former may be mitigated by RAG (retrieval-augmented generation) by anchoring answers in trusted clinical sources, there remains an open question as to whether the most effective model under this paradigm would be one that is fine-tuned on mental health data, or a more general and powerful model that succeeds purely on the basis of reasoning. In this paper, we perform a direct comparison by running four open-source models through the same RAG pipeline using ChromaDB: two generalist reasoners (Qwen2.5-3B and Phi-3-Mini) and two domain-specific fine-tunes (MentalHealthBot-7B and TherapyBot-7B). We use an LLM-as-a-Judge framework to automate evaluation over 50 turns. We find a clear trend: the generalist models outperform the domain-specific ones in empathy (3.72 vs. 3.26, $p < 0.001$) in spite of being much smaller (3B vs. 7B), and all models perform well in terms of safety, but the generalist models show better contextual understanding and are less prone to overfitting as we observe in the domain-specific models. Overall, our results indicate that for RAG-based therapy systems, strong reasoning is more important than training on mental health-specific vocabulary; i.e. a well-reasoned general model would provide more empathetic and balanced support than a larger narrowly fine-tuned model, so long as the answer is already grounded in clinical evidence.

BOISHOMMO: Holistic Approach for Bangla Hate Speech

One of the most alarming issues in digital society is hate speech (HS) on social media. The severity is so high that researchers across the globe are captivated by this domain. A notable amount of work has been conducted to address the identification and alarm system. However, a noticeable gap exists, especially for low-resource languages. Comprehensive datasets are the main problem among the constrained resource languages, such as Bangla. Interestingly, hate speech or any particular speech has no single dimensionality. Similarly, the hate component can simultaneously have multiple abusive attributes, which seems to be missed in the existing datasets. Thus, a multi-label Bangla hate speech dataset named BOISHOMMO has been compiled and evaluated in this work. That includes categories of HS across race, gender, religion, politics, and more. With over two thousand annotated examples, BOISHOMMO provides a nuanced understanding of hate speech in Bangla and highlights the complexities of processing non-Latin scripts. Apart from evaluating with multiple algorithmic approaches, it also highlights the complexities of processing Bangla text and assesses model performance. This unique multi-label approach enriches future hate speech detection and analysis studies for low-resource languages by providing a more nuanced, diverse dataset.