Mitigating Manifold Departure: Uncertainty-Aware Subspace Rectification for Trustworthy MLLM Decoding

The paper introduces Manifold-Guided Adaptive Projection (MGAP), a training-free decoding method designed to mitigate hallucinations in Multimodal Language Models (MLLMs) while preserving the model's semantic structure. MGAP utilizes singular value decomposition (SVD) to create a language-prior subspace and employs a consistency-aware gate to selectively attenuate the projected prior components during decoding. Experimental results on the POPE and CHAIR benchmarks demonstrate that MGAP significantly reduces hallucinations compared to existing methods, which is crucial for practitioners aiming to enhance the reliability of MLLM outputs in multimodal contexts.