New Framework Reduces Visual Hallucinations in MLLMs

Researchers have developed a novel framework aimed at reducing visual hallucinations and overconfident errors in multimodal large language models (MLLMs). MLLMs, despite their advanced capabilities in vision-language understanding, often produce unreliable outputs when visual evidence is weak, ambiguous, or inconsistent. Existing solutions typically focus on improving representation alignment or retrieval-augmented generation but lack mechanisms to quantify prediction reliability at an instance level. The proposed framework integrates retrieval-augmented reliability-aware inference. It constructs an external visual evidence database using pre-trained visual embeddings and nearest-neighbor retrieval. This retrieved evidence is then used to estimate prediction trustworthiness through several reliability indicators, including similarity strength, class-support agreement, and entropy-based uncertainty. Based on these signals, a decision gate determines whether the MLLM should confidently accept a prediction, answer with caution, or abstain/fallback when evidence is insufficient. A multimodal response-generation layer then crafts the final user-facing response, conditioned on this reliability decision. Experiments on ImageNet-100 demonstrated that this framework improved accepted prediction accuracy from 85.84% to 88.88% and reduced the hallucination-like accepted wrong-answer rate from 14.16% to 11.12%, all without requiring retraining of the core MLLM.