level: research
multimodal large language models often generate descriptions that conflict with visual inputs, a problem known as hallucination. this happens because the models rely too much on language patterns learned during training, which can override what they actually see. recent decoding methods try to fix this by penalizing these language priors, but they treat all priors as bad. in reality, some language priors are helpful and align with visual evidence, while others are not. blindly suppressing all priors can disrupt the model's internal representation structure, causing a drop in performance. the authors call this problem manifold departure.
to address this, the researchers developed manifold-guided adaptive projection, a geometry-aware decoding method that does not require retraining. it works by first building a subspace that captures language priors using singular value decomposition on hidden states from a blind run of the model. during decoding, each multimodal hidden state is projected onto this subspace. the key is an adaptive projection strength that depends on how uncertain the model is about the visual input. when visual evidence is strong, the projection is weak, preserving helpful priors. when visual evidence is weak, the projection is stronger, suppressing harmful priors.
experiments show that this method reduces hallucinations across several benchmarks without hurting performance on standard tasks. the adaptive mechanism helps maintain the model's semantic structure, avoiding the pitfalls of previous approaches. the method is lightweight and can be applied to existing models without any fine-tuning. it offers a practical way to make multimodal language models more trustworthy in applications like image captioning and visual question answering.
why it matters: this method improves the reliability of ai systems that combine vision and language, making them safer for real-world use where accurate descriptions of images are critical.