Enhanced medical diagnostic reasoning in small language models using reinforcement learning

Medical diagnosis is one of the most challenging cognitive tasks in healthcare, with diagnostic errors affecting millions of patients annually. Large Language Models (LLMs) have recently shown real potential in supporting clinical reasoning, however pre-trained models remain susceptible to cognitive biases, such as premature diagnostic closure, confirmation bias, and limited consideration of alternative hypotheses. These limitations are particularly pronounced in smaller models with fewer than 10 billion parameters. Most existing solutions heavily rely on prompting strategies or simply scaling model size, which does not fundamentally strengthen the underlying reasoning process. In this paper, we investigate whether diagnostic reasoning can be meaningfully improved in small language models through targeted reinforcement learning (RL) post-training. We adopt a two-stage training strategy: cold-start phase using supervised fine-tuning, followed by reinforcement learning with Group Relative Policy Optimization (GRPO). We combine this with a low-rank adaptation (LoRA) approach, which updates only a small fraction of model parameters for computational efficiency. We evaluate our approach on a 3B-parameter model (Qwen2.5-3B-Instruct) and an 8B-parameter model (Qwen3-8B-Base) using unstructured clinical notes from the MIMIC-IV dataset. The results demonstrate substantial improvements in diagnostic accuracy from 47.40% to 63.79% for the 3B model and from 49.54% to 66.56% for the 8B model, achieved with just 2554 training samples at a compute cost of approximately $3. These findings highlight that reinforcement learning-based post-training, combined with LoRA, offers a practical and cost-effective path to transforming small language models into capable diagnostic reasoners.