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

In three linesNew training-free MGAP method to reduce hallucinations in MLLMs. Constructs language-prior subspace via SVD and projects multimodal hidden states to selectively attenuate problematic components while preserving semantic structure. Improves POPE and CHAIR benchmarks.

Summary generated by Claude — human-verified