uncertainty-aware subspace rectification for trustworthy mllm decoding
a new training-free method reduces hallucinations in multimodal language models by adaptively projecting hidden states onto a language-prior subspace, preserving helpful priors while suppressing harmful ones.