TIF-GRPO: New RL method stops medical AI from hallucinating diagnoses

A team of researchers (Tianwei Lin, Zhongwei Qiu, Jie Cao, et al.) has proposed a novel reinforcement learning framework called Trajectory-Integral Feedback GRPO (TIF-GRPO) to tackle a critical flaw in medical vision-language models (VLMs) used for 3D Computed Tomography (CT) analysis. Current RL paradigms rely on lexical proxy signals that cause 'Evaluation Hallucinations'—models optimize linguistic fluency rather than factual clinical correctness, leading to diagnostically dangerous errors. The team introduces the Clinical Abnormality Benchmarking Substrate (CABS), a structured system that decomposes radiology reports into verifiable clinical semantic units, revealing a 'Mechanistic Divergence' where surface-similarity rewards bypass medical facts.

TIF-GRPO applies control-theoretic principles by formulating clinical reasoning as a pseudo-temporal trajectory for anomaly discovery. It regulates anatomy-aware rewards through an integral feedback loop that penalizes persistent omissions as cumulative state errors and suppresses hallucinations as excessive control effort. Experiments on 3D CT benchmarks demonstrated significant improvements in abnormality detection and clinical faithfulness, establishing a new paradigm for fine-grained regulation in medical VLMs. The project is available on GitHub under arXiv:2605.20277.