Title: OncoReason: Enhancing Clinical Reasoning in Large Language Models for Reliable and Interpretable Survival Prognosis

Abstract:

Accurate and transparent prediction of cancer treatment outcomes is essential, especially when dealing with complex and varied clinical information. Although large language models (LLMs) have demonstrated significant proficiency in biomedical natural language processing, they frequently lack the structured reasoning abilities necessary for critical decision-making support. To address this gap, we introduce a comprehensive, multi-task learning framework designed to integrate clinical reasoning into autoregressive LLMs for outcome prediction using the MSK-CHORD dataset. This approach trains models to simultaneously execute binary survival classification, regression for continuous survival times, and the generation of natural language rationales.

We assess three distinct alignment methodologies: (1) conventional supervised fine-tuning (SFT); (2) SFT augmented with Chain-of-Thought (CoT) prompting to encourage step-by-step logical deduction; and (3) Group Relative Policy Optimization (GRPO), a reinforcement learning technique that aligns model outputs with expert-derived reasoning paths.

Evaluations conducted using LLaMa3-8B and Med42-8B backbone architectures reveal that CoT prompting boosts the F1 score by 6.0 points and decreases the Mean Absolute Error (MAE) by 12%. Meanwhile, GRPO delivers state-of-the-art results in both predictive accuracy and interpretability, as measured by BLEU, ROUGE, and BERTScore metrics. Additionally, our analysis indicates that current biomedical LLMs often struggle to generate valid reasoning traces due to inherent architectural limitations. These results highlight the necessity of reasoning-aware alignment in multi-task clinical modeling and establish a new standard for trustworthy, interpretable LLMs in the field of precision oncology.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC