New framework cuts medical AI hallucinations by 35% without retraining

Multimodal large language models (MLLMs) show strong promise for medical tasks like clinical VQA and radiology report generation. However, inference-time hallucinations remain a critical barrier—models can produce fluent but inaccurate conclusions that contradict imaging evidence. Existing mitigation strategies rely on additional training, external retrieval, or multi-stage verification, increasing cost and complexity while often failing to generalize. To address this, researchers from multiple institutions propose SPRG (Synergistic Perception-Reasoning Governance), a holistic, training-free framework that systematically reduces hallucinations through dual-side evidence injection. The method uses MedSAM to extract region-of-interest (ROI) priors, then recalibrates the visual perception trajectory via ROI-guided activation modulation. It also anchors the textual reasoning trajectory by mapping anatomical coordinates into discrete semantic tokens as verifiable external memory. A task-aware dynamic router selects modality-specific interventions based on task semantics, balancing perceptual grounding and linguistic fluency.

SPRG was systematically evaluated on 2 tasks and 5 datasets using multiple models: LLaVA-1.5-7B, LLaVA-Med-1.5-7B, Qwen3-VL-8B/32B, and InternVL-3.5-8B/38B. Controlled ablations and visualizations confirmed that SPRG consistently outperforms baselines across medical benchmarks, achieving up to ~6% improvement in close-ended accuracy and a ~35% reduction in open-ended hallucinations. The paper has been accepted by MICCAI 2026 as an Early Accept (Top 9%). The code is available on GitHub, enabling researchers and practitioners to integrate this method without retraining or external databases. This work represents a significant step toward trustworthy deployment of medical AI, particularly in clinical settings where hallucination risk is unacceptable.