Researchers from Tsinghua University have identified a form of object hallucination in multimodal large models (MLLMs) that they say is rooted in visual feature extraction rather than language priors alone. In a paper at ACM MM 2026 (the 34th ACM International Conference on Multimedia), the team proposed that when model outputs are very short—such as answering only "Yes" or "No"—reasoning relies more heavily on the visual modality, revealing what they named visual-origin hallucination. They proposed a corresponding solution, ACFT. Experiments show ACFT achieves excellent performance on the POPE, MME, and four description-level hallucination benchmarks across the LLaVA, MiniGPT-4, and Qwen2.5-VL models, using only 0.9% of the COCO dataset and without adding any inference overhead. The article notes object hallucination has long been a core challenge to MLLM reliability, with existing work largely following the "language prior" line of inquiry and mitigation methods categorized by intervention point, including input-level interventions during decoding, external grounding modules for post-output processing, and post-training alignment.